Many Thumb1 instructions are defined to set CPSR if executed outside an IT
block, but leave it alone from inside one. In MachineIR this is represented by
whether an optional register is CPSR or NoReg (0), and affects how the
instructions are printed.
This sets the instruction to the appropriate form during if-conversion.
Added extra patterns to VABD instruction so it is selected in place of VSUB and VABS. Added corresponding regression test too.
Differential Revision: https://reviews.llvm.org/D84500
Similar to 8fa824d7a3 but this time for MLA patterns, this selects
predicated vmlav/vmlava/vmlalv/vmlava instructions from
vecreduce.add(select(p, mul(x, y), 0)) nodes.
Differential Revision: https://reviews.llvm.org/D84102
Given a vecreduce.add(select(p, x, 0)), we can convert that to a
predicated vaddv, as the else value for the select is the identity
value, a zero. That is what this patch does for the vaddv, vaddva,
vaddlv and vaddlva instructions, copying the existing patterns to also
handle predication through a select.
Differential Revision: https://reviews.llvm.org/D84101
For a long time, the InstCombine pass handled target specific
intrinsics. Having target specific code in general passes was noted as
an area for improvement for a long time.
D81728 moves most target specific code out of the InstCombine pass.
Applying the target specific combinations in an extra pass would
probably result in inferior optimizations compared to the current
fixed-point iteration, therefore the InstCombine pass resorts to newly
introduced functions in the TargetTransformInfo when it encounters
unknown intrinsics.
The patch should not have any effect on generated code (under the
assumption that code never uses intrinsics from a foreign target).
This introduces three new functions:
TargetTransformInfo::instCombineIntrinsic
TargetTransformInfo::simplifyDemandedUseBitsIntrinsic
TargetTransformInfo::simplifyDemandedVectorEltsIntrinsic
A few target specific parts are left in the InstCombine folder, where
it makes sense to share code. The largest left-over part in
InstCombineCalls.cpp is the code shared between arm and aarch64.
This allows to move about 3000 lines out from InstCombine to the targets.
Differential Revision: https://reviews.llvm.org/D81728
This is very similar to 243970d03cace2, but handling a slightly
different form of predicated operations. When starting with a pattern of
the form select(p, BinOp(x, y), x), Instcombine will often transform
this to BinOp(x, select(p, y, 0)), where 0 is the identity value of the
binop (0 for adds/subs, 1 for muls, -1 for ands etc). This adds the
patterns that transforms those back into predicated binary operations.
There is also a very minor adjustment to tablegen null_frag in here, to
allow it to also be recognized as a PatLeaf node, so that it can be used
in MVE_TwoOpPattern to easily exclude the cases where we do not need the
alternate transform.
Differential Revision: https://reviews.llvm.org/D84091
Most MVE instructions can be predicated to fold a select into the
instruction, using the predicate and the selects else as a passthough.
This adds tablegen patterns for most two operand instructions using the
newly added TwoOpPattern from 1030e82598.
Differential Revision: https://reviews.llvm.org/D83222
These extra vcvt instructions were missed from 74ca67c109 because they
live in a different Domain, but should be treated in the same way.
Differential Revision: https://reviews.llvm.org/D83204
As far as I can tell, it should not be necessary for VCTP to be
unpredictable in tail predicated loops. Either it has a a valid loop
counter as a operand which will naturally keep it in the right loop, or
it doesn't and it won't be converted to a tail predicated loop. Not
marking it as having side effects allows it to be scheduled more cleanly
for cases where it is not expected to become a tail predicate loop.
Differential Revision: https://reviews.llvm.org/D83907
There were cases where a do-while loop would be converted to a while
loop before finding out that it would be unsafe to expand the SCEV in
this situation and then bailing out of hardware loop conversion.
This patch checks if it would be unsafe to expand the SCEV and if so stops converting the do-while into a while, allowing conversion to a hardware loop.
Differential Revision: https://reviews.llvm.org/D83953
In an upcoming AMDGPU patch, the scalar cases will be legal and vector
ops should be scalarized, rather than producing a long sequence of
vector ops which will also need to be scalarized.
Use a lazy heuristic that seems to work and improves the thumb2 MVE
test.
Vector bitwise selects are matched by pseudo VBSP instruction
and expanded to VBSL/VBIT/VBIF after register allocation
depend on operands registers to minimize extra copies.
This adds a peephole optimisation to turn a t2MOVccr that could not be
folded into any other instruction into a CSEL on 8.1-m. The t2MOVccr
would usually be expanded into a conditional mov, that becomes an IT;
MOV pair. We can instead generate a CSEL instruction, which can
potentially be smaller and allows better register allocation freedom,
which can help reduce codesize. Performance is more variable and may
depend on the micrarchitecture details, but initial results look good.
If we need to control this per-cpu, we can add a subtarget feature as we
need it.
Original patch by David Penry.
Differential Revision: https://reviews.llvm.org/D83566
This refactors option -disable-mve-tail-predication to take different arguments
so that we have 1 option to control tail-predication rather than several
different ones.
This is also a prep step for D82953, in which we want to reject reductions
unless that is requested with this option.
Differential Revision: https://reviews.llvm.org/D83133
Whether an instruction is deemed to have side effects in determined by
whether it has a tblgen pattern that emits a single instruction.
Because of the way a lot of the the vcvt instructions are specified
either in dagtodag code or with patterns that emit multiple
instructions, they don't get marked as not having side effects.
This just marks them as not having side effects manually. It can help
especially with instruction scheduling, to not create artificial
barriers, but one of these tests also managed to produce fewer
instructions.
Differential Revision: https://reviews.llvm.org/D81639
Given a loop with two subloops, it should be possible for both to be
converted to hardware loops. That's what this patch does, simply enough.
It slightly alters the loop iterating order to try and convert all
subloops. If one (or more) succeeds, it stops as before.
Differential Revision: https://reviews.llvm.org/D78502
Move the Thumb2SizeReduce pass to before IfConversion when optimising
for minimal code size.
Running the Thumb2SizeReduction pass before IfConversionallows T1
instructions to propagate to the final output, rather than the
ifConverter modifying T2 instructions and preventing them from being
reduced later.
This change does introduce a regression regarding execution time, so
it's only applied when optimising for size.
Running the LLVM Test Suite with this change produces a geomean
difference of -0.1% for the size..text metric.
Differential Revision: https://reviews.llvm.org/D82439
While validating live-out values, record instructions that look like
a reduction. This will comprise of a vector op (for now only vadd),
a vorr (vmov) which store the previous value of vadd and then a vpsel
in the exit block which is predicated upon a vctp. This vctp will
combine the last two iterations using the vmov and vadd into a vector
which can then be consumed by a vaddv.
Once we have determined that it's safe to perform tail-predication,
we need to change this sequence of instructions so that the
predication doesn't produce incorrect code. This involves changing
the register allocation of the vadd so it updates itself and the
predication on the final iteration will not update the falsely
predicated lanes. This mimics what the vmov, vctp and vpsel do and
so we then don't need any of those instructions.
Differential Revision: https://reviews.llvm.org/D75533
After the rewrite of this pass (D79175) I missed one thing: the inserted VCTP
intrinsic can be cloned to exit blocks if there are instructions present in it
that perform the same operation, but this wasn't triggering anymore. However,
it turns out that for handling reductions, see D75533, it's actually easier not
not to have the VCTP in exit blocks, so this removes that code.
This was possible because it turned out that some other code that depended on
this, rematerialization of the trip count enabling more dead code removal
later, wasn't doing much anymore due to more aggressive dead code removal that
was added to the low-overhead loops pass.
Differential Revision: https://reviews.llvm.org/D82773
This patch stops the trunc, rint, round, floor and ceil intrinsics from blocking tail predication.
Differential Revision: https://reviews.llvm.org/D82553
MVE has native reductions for integer add and min/max. The others need
to be expanded to a series of extract's and scalar operators to reduce
the vector into a single scalar. The default codegen for that expands
the reduction into a series of in-order operations.
This modifies that to something more suitable for MVE. The basic idea is
to use vector operations until there are 4 remaining items then switch
to pairwise operations. For example a v8f16 fadd reduction would become:
Y = VREV X
Z = ADD(X, Y)
z0 = Z[0] + Z[1]
z1 = Z[2] + Z[3]
return z0 + z1
The awkwardness (there is always some) comes in from something like a
v4f16, which is first legalized by adding identity values to the extra
lanes of the reduction, and which can then not be optimized away through
the vrev; fadd combo, the inserts remain. I've made sure they custom
lower so that we can produce the pairwise additions before the extra
values are added.
Differential Revision: https://reviews.llvm.org/D81397
Similar to the recent patch for fpext, this adds vcvtb and vcvtt with
insert into vector instruction selection patterns for fptruncs. This
helps clear up a lot of register shuffling that we would otherwise do.
Differential Revision: https://reviews.llvm.org/D81637
We current extract and convert from a top lane of a f16 vector using a
VMOVX;VCVTB pair. We can simplify that to use a single VCVTT. The
pattern is mostly copied from a vector extract pattern, but produces a
VCVTTHS f32 directly.
This had to move some code around so that ARMInstrVFP had access to the
required pattern frags that were previously part of ARMInstrNEON.
Differential Revision: https://reviews.llvm.org/D81556
This lowers intrinsic @llvm.get.active.lane.mask to a setcc node, i.e. an icmp
ule, and creates vectors for its 2 arguments on which the comparison is
performed.
Differential Revision: https://reviews.llvm.org/D82292
This extends PerformSplittingToWideningLoad to also handle FP_Ext, as
well as sign and zero extends. It uses an integer extending load
followed by a VCVTL on the bottom lanes to efficiently perform an fpext
on a smaller than legal type.
The existing code had to be rewritten a little to not just split the
node in two and let legalization handle it from there, but to actually
split into legal chunks.
Differential Revision: https://reviews.llvm.org/D81340
This adds code to lower f16 to f32 fp_exts's using an MVE VCVT
instructions, similar to a recent similar patch for fp_trunc. Again it
goes through the lowering of a BUILD_VECTOR, but is slightly simpler
only having to deal with interleaved indices. It adds a VCVTL node to
lower to, similar to VCVTN.
Differential Revision: https://reviews.llvm.org/D81339
This splits MVE vector stores of a fp_trunc in the same way that we do
for standard trunc's. It extends PerformSplittingToNarrowingStores to
handle fp_round, splitting the store into pieces and adding a VCVTNb to
perform the actual fp_round. The actual store is then converted to an
integer store so that it can truncate bottom lanes of the result.
Differential Revision: https://reviews.llvm.org/D81141
This adds code to lower f32 to f16 fp_trunc's using a pair of MVE VCVT
instructions. Due to v4f16 not being legal, fp_round are often split up
fairly early. So this reconstructs the vcvt's from a buildvector of
fp_rounds from two vector inputs. Something like:
BUILDVECTOR(FP_ROUND(EXTRACT_ELT(X, 0),
FP_ROUND(EXTRACT_ELT(Y, 0),
FP_ROUND(EXTRACT_ELT(X, 1),
FP_ROUND(EXTRACT_ELT(Y, 1), ...)
It adds a VCVTN node to handle this, which like VMOVN or VQMOVN lowers
into the top/bottom lanes of an MVE instruction.
Differential Revision: https://reviews.llvm.org/D81139
LDRD and STRD along with UBFX and SBFX are selected from DAGToDAG
transforms, so do not have tblgen patterns. They don't get marked as
having side effects so cannot be scheduled as efficiently as you would
like.
This specifically marks then as not having side effects.
Differential Revision: https://reviews.llvm.org/D82358
Summary:
Half-precision floating point arguments and returns are currently
promoted to either float or int32 in clang's CodeGen and there's
no existing support for the lowering of `half` arguments and returns
from IR in AArch32's backend.
Such frontend coercions, implemented as coercion through memory
in clang, can cause a series of issues in argument lowering, as causing
arguments to be stored on the wrong bits on big-endian architectures
and incurring in missing overflow detections in the return of certain
functions.
This patch introduces the handling of half-precision arguments and returns in
the backend using the actual "half" type on the IR. Using the "half"
type the backend is able to properly enforce the AAPCS' directions for
those arguments, making sure they are stored on the proper bits of the
registers and performing the necessary floating point convertions.
Reviewers: rjmccall, olista01, asl, efriedma, ostannard, SjoerdMeijer
Reviewed By: ostannard
Subscribers: stuij, hiraditya, dmgreen, llvm-commits, chill, dnsampaio, danielkiss, kristof.beyls, cfe-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D75169
To set up a tail-predicated loop, we need to to calculate the number of
elements processed by the loop. We can now use intrinsic
@llvm.get.active.lane.mask() to do this, which is emitted by the vectoriser in
D79100. This intrinsic generates a predicate for the masked loads/stores, and
consumes the Backedge Taken Count (BTC) as its second argument. We can now use
that to reconstruct the loop tripcount, instead of the IR pattern match
approach we were using before.
Many thanks to Eli Friedman and Sam Parker for all their help with this work.
This also adds overflow checks for the different, new expressions that we
create: the loop tripcount, and the sub expression that calculates the
remaining elements to be processed. For the latter, SCEV is not able to
calculate precise enough bounds, so we work around that at the moment, but is
not entirely correct yet, it's conservative. The overflow checks can be
overruled with a force flag, which is thus potentially unsafe (but not really
because the vectoriser is the only place where this intrinsic is emitted at the
moment). It's also good to mention that the tail-predication pass is not yet
enabled by default. We will follow up to see if we can implement these
overflow checks better, either by a change in SCEV or we may want revise the
definition of llvm.get.active.lane.mask.
Differential Revision: https://reviews.llvm.org/D79175
In more complicated loops we can easily hit the complexity limits of
loop strength reduction. If we do and filtering occurs, it's all too
easy to remove the wrong formulae for post-inc preferring accesses due
to it attempting to maximise register re-use. The patch adds an
alternative filtering step when the target is preferring postinc to pick
postinc formulae instead, hopefully lowering the complexity to below the
limit so that aggressive filtering is not needed.
There is also a change in here to stop considering existing addrecs as
free under postinc. We should already be modelling them as a reg so
don't want it to cause us to get the cost wrong. (I'm not sure that code
makes sense in general, but there are X86 tests specifically for it
where it seems to be helping so have left it around for the standard
non-post-inc case).
Differential Revision: https://reviews.llvm.org/D80273
These code patterns attempt to call isVMOVModifiedImm on a splat of i1
values, leading to an unreachable being hit. I've guarded the call on a
more specific set of sizes, as i1 vectors are legal under MVE.
Differential Revision: https://reviews.llvm.org/D81860
Similar to VMOVN, a VQMOVN will only demand the top/bottom lanes of it's
first input. However unlike VMOVN it will need access to the entire
second argument, as that value is saturated not just moved in place.
Differential Revision: https://reviews.llvm.org/D80515
These patterns for i8 and i16 VMLA's were missing. They end up from
legalized vector.reduce.add.v8i16 and vector.reduce.add.v16i8, and
although the instruction works differently (the mul and add are
performed in a higher precision), I believe it is OK because only an
i8/i16 are demanded from them, and so the results will be the same. At
least, they pass any testing I can think to run on them.
There are some tests that end up looking worse, but are quite artificial
due to passing half vector types through a call boundary. I would not
expect the vmull to realistically come up like that, and a vmlava is
likely better a lot of the time.
Differential Revision: https://reviews.llvm.org/D80524
Summary:
In Thumb2's frame index rewriting process, the address mode i8s4, which
is used by LDRD and STRD instructions, is handled by taking the
immediate offset operand and multiplying it by 4.
This behaviour is wrong, however. In this specific address mode, the
MachineInstr's immediate operand is already in the expected form. By
consequence of that, multiplying it once more by 4 yields a flawed
offset value, four times greater than it should be.
Differential Revision: https://reviews.llvm.org/D80557
This temporarily reverts commit 7019cea26d.
It seems that, for some targets, there are instructions with a lot of memory operands (probably more than would be expected). This causes a lot of buildbots to timeout and notify failed builds. While investigations are ongoing to find out why this happens, revert the changes.
Summary:
To support all targets, the mayAlias member function needs to support instructions with multiple operands.
This revision also changes the order of the emitted instructions in some test cases.
Reviewers: efriedma, hfinkel, craig.topper, dmgreen
Reviewed By: efriedma
Subscribers: MatzeB, dmgreen, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D80161
This is split off from D80316, slightly tightening the definition of overloaded
hardwareloop intrinsic llvm.loop.decrement.reg specifying that both operands
its result have the same type.
Previously, the LowOverheadLoops pass couldn't handle VPT blocks
with conditions, or with multiple VCTPs. This patch improves the
LowOverheadLoops pass so it can handle those cases.
It also adds support for VCMPs before the VCTP.
Differential Revision: https://reviews.llvm.org/D78206
Given a VQMOVN(VSHR), we can fold that into a VQSHRN simply enough using
a few tablegen patterns.
Differential Revision: https://reviews.llvm.org/D77720
This adds two combines for VMOVN, one to fold
VMOVN[tb](c, VQMOVNb(a, b)) => VQMOVN[tb](c, b)
The other to perform demand bits analysis on the lanes of a VMOVN. We
know that only the bottom lanes of the second operand and the top or
bottom lanes of the Qd operand are needed in the result, depending on if
the VMOVN is bottom or top.
Differential Revision: https://reviews.llvm.org/D77718
This adds some custom lowering for VQMOVN, an instruction that can be
used to perform saturating truncates from a pair of min(max(X, -0x8000),
0x7fff), providing those constants are correct. This leaves a VQMOVNBs
which saturates the value and inserts that into the bottom lanes of an
existing vector. We then need to do something with the other lanes,
extending the value using a vmovlb.
Ideally, as will often be the case, only the bottom lane of what remains
will be demanded, allowing the vmovlb to be removed. Which should mean
the instruction is either equal or a win most of the time, and allows
some extra follow-up folding to happen.
Differential Revision: https://reviews.llvm.org/D77590
Adds support to build pre-incrementing scatters.
If the increment (i.e., add instruction) that is merged into
the scatter is the loop increment, an incrementing write-back
scatter can be built, which then assumes the role of the loop
increment.
Differential Revision: https://reviews.llvm.org/D79859
Under MVE a vdup will always take a gpr register, not a floating point
value. During DAG combine we convert the types to a bitcast to an
integer in an attempt to fold the bitcast into other instructions. This
is OK, but only works inside the same basic block. To do the same trick
across a basic block boundary we need to convert the type in
codegenprepare, before the splat is sunk into the loop.
This adds a convertSplatType function to codegenprepare to do that,
putting bitcasts around the splat to force the type to an integer. There
is then some adjustment to the code in shouldSinkOperands to handle the
extra bitcasts.
Differential Revision: https://reviews.llvm.org/D78728
Similar to fmul/fadd, we can sink a splat into a loop containing a fma
in order to use more register instruction variants. For that there are
also adjustments to the sinking code to handle more than 2 arguments.
Differential Revision: https://reviews.llvm.org/D78386
This patch adds a new TTI hook to allow targets to tell LSR that
a chain including some instruction is already profitable and
should not be optimized. This patch also adds an implementation
of this TTI hook for ARM so LSR doesn't optimize chains that include
the VCTP intrinsic.
Differential Revision: https://reviews.llvm.org/D79418
allocas in LLVM IR have a specified alignment. When that alignment is
specified, the alloca has at least that alignment at runtime.
If the specified type of the alloca has a higher preferred alignment,
SelectionDAG currently ignores that specified alignment, and increases
the alignment. It does this even if it would trigger stack realignment.
I don't think this makes sense, so this patch changes that.
I was looking into this for SVE in particular: for SVE, overaligning
vscale'ed types is extra expensive because it requires realigning the
stack multiple times, or using dynamic allocation. (This currently isn't
implemented.)
I updated the expected assembly for a couple tests; in particular, for
arg-copy-elide.ll, the optimization in question does not increase the
alignment the way SelectionDAG normally would. For the rest, I just
increased the specified alignment on the allocas to match what
SelectionDAG was inferring.
Differential Revision: https://reviews.llvm.org/D79532
Unlike Neon, MVE does not have a way of duplicating from a vector lane,
so a VDUPLANE currently selects to a VDUP(move_from_lane(..)). This
forces that to be done earlier as a dag combine to allow other folds to
happen.
It converts to a VDUP(EXTRACT). On FP16 this is then folded to a
VGETLANEu to prevent it from creating a vmovx;vmovhr pair, using a
single move_from_reg instead.
Differential Revision: https://reviews.llvm.org/D79606
If the SimplifyMultipleUseDemandedBits calls BITCASTs that peek through back to the original type then we can remove the BITCASTs entirely.
Differential Revision: https://reviews.llvm.org/D79572
CorrectExtraCFGEdges function.
The latter was a workaround for "Various pieces of code" leaving bogus
extra CFG edges in place. Where by "various" it meant only
IfConverter::MergeBlocks, which failed to clear all of the successors
of dead blocks it emptied out. This wouldn't matter a whole lot,
except that the dead blocks remained listed as predecessors of
still-useful blocks, inhibiting optimizations.
This fix slightly changed two thumb tests, because the correct CFG
successors allowed for the "diamond" if-conversion pattern to be
detected, when it could only use "simple" before.
Additionally, the removal of a now-redundant call to analyzeBranch
(with AllowModify=true) in BranchFolder::OptimizeFunction caused a
later check for an empty block in BranchFolder::OptimizeBlock to
fail. Correct this by moving the call to analyzeBranch in
OptimizeBlock higher.
Differential Revision: https://reviews.llvm.org/D79527
Enables the MVEGatherScatterLowering pass to build
pre-incrementing gathers. Incrementing writeback gathers
are built when it is possible to replace the loop increment
instruction.
Differential Revision: https://reviews.llvm.org/D76786
Much like the similar combine added recently for VMOVrh load, this
adds a fold for VMOVhr load turning it into a vldr.f16 as opposed to a
vldrh and vmov.f16.
Differential Revision: https://reviews.llvm.org/D78714
If we get into the situation where we are extracting from a VDUP, the
extracted value is just the origin, so long as the types match or we can
bitcast between the two.
Differential Revision: https://reviews.llvm.org/D78708
The idea, under MVE, is to introduce more bitcasts around VDUP's in an
attempt to get the type correct across basic block boundaries. In order
to do that without other regressions we need a few fixups, of which this
is the first. If the code is a bitcast of a VDUP, we can convert that
straight into a VDUP of the new type, so long as they have the same
size.
Differential Revision: https://reviews.llvm.org/D78706
LSR has some logic that tries to aggressively reuse registers in
formula. This can lead to sub-optimal decision in complex loops where
the backend it trying to use shouldFavorPostInc. This disables the
re-use in those situations.
Differential Revision: https://reviews.llvm.org/D79301
A PREDICATE_CAST(PREDICATE_CAST(X)) can be converted to a
PREDICATE_CAST(X) as the operation can convert between any forms of
predicates (v4i1/v8i1/v16i1/i32). Unfortunately I got the type wrong on
one of the rarer converts, which would lead to invalid nodes during
isel. This fixes it up to use the correct type.
Differential Revision: https://reviews.llvm.org/D79402
This patch makes the folding of or(A, B) into not(and(not(A), not(B)))
more agressive for I1 vector. This only affects Thumb2 MVE and improves
codegen, because it removes a lot of msr/mrs instructions on VPR.P0.
This patch also adds a xor(vcmp) -> !vcmp fold for MVE.
Differential Revision: https://reviews.llvm.org/D77202
This patch adds an implementation of PerformVSELECTCombine in the
ARM DAG Combiner that transforms vselect(not(cond), lhs, rhs) into
vselect(cond, rhs, lhs).
Normally, this should be done by the target-independent DAG Combiner,
but it doesn't handle the kind of constants that we generate, so we
have to reimplement it here.
Differential Revision: https://reviews.llvm.org/D77712
The two code paths have the same goal, legalizing a load of a non-byte-sized vector by loading the "flattened" representation in memory, slicing off each single element and then building a vector out of those pieces.
The technique employed by `ExpandLoad` is slightly more convoluted and produces slightly better codegen on ARM, AMDGPU and x86 but suffers from some bugs (D78480) and is wrong for BE machines.
Differential Revision: https://reviews.llvm.org/D79096
This changes the logic with lowering fp16 bitcasts to always produce
either a VMOVhr or a VMOVrh, instead of only trying to do it with
certain surrounding nodes. To perform the same optimisations demand bits
and known bits information has been added for them.
Differential Revision: https://reviews.llvm.org/D78587
There are some intrinsics like this that currently block tail
predication, but should be fine. This allows fma through, as the one
that I ran into. There may be others that need the same treatment but
I've only done this one here.
Differential Revision: https://reviews.llvm.org/D78385
This adds some extra processing into the Pre-RA ARM load/store optimizer
to detect and merge MVE loads/stores and adds of the same base. This we
don't always turn into a post-inc during ISel, and due to the nature of
it being a graph we don't always know an order to use for the nodes, not
knowing which nodes to make post-inc and which to use the new post-inc
of. After ISel, we have an order that we can use to post-inc the
following instructions.
So this looks for a loads/store with a starting offset of 0, and an
add/sub from the same base, plus a number of other loads/stores. We then
do some checks and convert the zero offset load/store into a postinc
variant. Any loads/stores after it have the offset subtracted from their
immediates. For example:
LDR #4 LDR #4
LDR #0 LDR_POSTINC #16
LDR #8 LDR #-8
LDR #12 LDR #-4
ADD #16
It only handles MVE loads/stores at the moment. Normal loads/store will
be added in a followup patch, they just have some extra details to
ensure that we keep generating LDRD/LDM successfully.
Differential Revision: https://reviews.llvm.org/D77813
Add patterns that use a normal, non-wrapping, add and sub nodes along
with an arm vshr imm node.
Differential Revision: https://reviews.llvm.org/D77065
Fix for the address optimization for gathers and scatters which would in
some complex cases push out instructions not to the vector loop preheader,
but to other locations as well which lead to a scrambled order and the
compilation failing.
This patch ensures that said instructions are always pushed to the end
of the vector loop preheader.
Differential Revision: https://reviews.llvm.org/D78293
Summary:
The INLINEASM MIR instructions use immediate operands to encode the values of some operands.
The MachineInstr pretty printer function already handles those operands and prints human readable annotations instead of the immediates. This patch adds similar annotations to the output of the MIRPrinter, however uses the new MIROperandComment feature.
Reviewers: SjoerdMeijer, arsenm, efriedma
Reviewed By: arsenm
Subscribers: qcolombet, sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78088
The pass was incorrectly reverting back to a "T" when something wrote
to VPR inside a "E" block. This is not the correct behaviour, the
predicate should stay the same.
Differential Revision: https://reviews.llvm.org/D77798
This patch adds an analysis of the offset addresses used by gathers
and scatters to the MVEGatherScatterLowering pass to find
multiplications and additions that are loop invariant and thus can
be moved into the loop preheader, avoiding to execute them each time.
Differential Revision: https://reviews.llvm.org/D76681
If the stack pointer is altered for local variables and we are generating
Thumb2 execute-only code the .pad directive is missing.
Usually the size of the adjustment is stored in a PC-relative location
and loaded into a register which is then added to the stack pointer.
However when we are generating execute-only code code the size of the
adjustment is instead generated using the MOVW/MOVT instruction pair.
As a by product of handling the execute-only case this also fixes an
existing issue that in the none execute-only case the .pad directive was
generated against the load of the constant to a register instruction,
instead of the instruction which adds the register to the stack pointer.
Differential Revision: https://reviews.llvm.org/D76849
Currently when the target is big-endian vmov.i64 reverses the order of the two
words of the vector. This is correct only when the underlying element type is
32-bit, as actually what it should be doing is considering it a vector of the
underlying type and reversing the elements of that.
Differential Revision: https://reviews.llvm.org/D76515
If we have an element-wise vmov immediate instruction then a subsequent vrev
with width greater or equal to the vmov element width, then that vrev won't do
anything. Add a DAG combine to convert bitcasts that would become such vrevs
into vector_reg_casts instead.
Differential Revision: https://reviews.llvm.org/D76514
This adds MVE vmull patterns, which are conceptually the same as
mul(vmovl, vmovl), and so the tablegen patterns follow the same
structure.
For i8 and i16 this is simple enough, but in the i32 version the
multiply (in 64bits) is illegal, meaning we need to catch the pattern
earlier in a dag fold. Because bitcasts are involved in the zext
versions and the patterns are a little different in little and big
endian. I have only added little endian support in this patch.
Differential Revision: https://reviews.llvm.org/D76740
The unpredictable/hasSideEffects flag is usually inferred by tablegen
from whether the instruction has a tablegen pattern (and that pattern
only has a single output instruction). Now that the MVE intrinsics are
all committed and producing code, the remaining instructions still
marked as unpredictable need to be specially handled. This adds the flag
directly to instructions that need it, notably the V*MLAL instructions
and some of the MOV's.
Differential Revision: https://reviews.llvm.org/D76910
This allows the MVE VPT Block insertion pass to remove VPNOTs in
order to create more complex VPT blocks such as TE, TEET, TETE, etc.
Differential Revision: https://reviews.llvm.org/D75993
Add a bit more logic into the 'FalseLaneZeros' tracking to enable
horizontal reductions and also make the VADDV variants
validForTailPredication.
Differential Revision: https://reviews.llvm.org/D76708
In the original batch of MVE VMOVimm code generation VMOV.i64 was left
out due to the way it was done downstream. It turns out that it's fairly
simple though. This adds the codegen for it, similar to NEON.
Bigendian is technically incorrect in this version, which John is fixing
in a Neon patch.
Given that some instructions generate wider result elements than
their inputs, flag them as being able to generate non zeros in the
false lanes.
Differential Revision: https://reviews.llvm.org/D76766
Some MVE floating point instructions have gpr register variants that take
the scalar gpr value and splat them to all lanes. In order to accept
them in loops, the shuffle_vector and insert need to be sunk down into
the loop, next to the instruction so that ISel can see the whole
pattern.
This does that sinking for FAdd, FSub, FMul and FCmp. The patterns for
mul are slightly more constrained as there are no fms variants taking
register arguments.
Differential Revision: https://reviews.llvm.org/D76023
Summary:
This patch implements the following CDE intrinsics:
T __arm_vcx1q_m(int coproc, T inactive, uint32_t imm, mve_pred_t p);
T __arm_vcx2q_m(int coproc, T inactive, U n, uint32_t imm, mve_pred_t p);
T __arm_vcx3q_m(int coproc, T inactive, U n, V m, uint32_t imm, mve_pred_t p);
T __arm_vcx1qa_m(int coproc, T acc, uint32_t imm, mve_pred_t p);
T __arm_vcx2qa_m(int coproc, T acc, U n, uint32_t imm, mve_pred_t p);
T __arm_vcx3qa_m(int coproc, T acc, U n, V m, uint32_t imm, mve_pred_t p);
The intrinsics are not part of the released ACLE spec, but internally at
Arm we have reached consensus to add them to the next ACLE release.
Reviewers: simon_tatham, MarkMurrayARM, ostannard, dmgreen
Reviewed By: simon_tatham
Subscribers: kristof.beyls, hiraditya, danielkiss, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76610
Summary:
These were accidentally left out of D76123. I added tests for the
other three instructions in this small cross-product family (vqdmlah,
vqrdmlah, vqrdmlash) but missed this one.
Reviewers: miyuki
Reviewed By: miyuki
Subscribers: kristof.beyls, dmgreen, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76714
This adds a simple fold to combine VMOVrh load to a integer load.
Similar to what is already performed for BITCAST, but needs to account
for the types being of different sizes, creating an zero extending load.
Differential Revision: https://reviews.llvm.org/D76485
We deliberately split stores of the form
store(truncate(larger-than-legal-type)) into two stores, allowing each
store to perform part of the truncate for free.
There are times however where it makes more sense to use VMOVN to
de-interlace the results back into a single vector, and store that in
one go. This adds a check for that situation, not splitting the store if
it looks like a VMOVN can be more useful.
Differential Revision: https://reviews.llvm.org/D76511
When decided whether to generate a post-inc load/store, look at the
other memory nodes that use the same base address and, if any proceed
the current node, then don't do the combine.
The change only seems to be affecting the Arm backend, which I was
surprised at, but it appears to fix a lot of our issues around MVE
masked load/stores having to store a temporary address after an early
post-increment on a shared base address.
Differential Revision: https://reviews.llvm.org/D75847
Summary:
I've implemented them as target-specific IR intrinsics rather than
using `@llvm.experimental.vector.reduce.add`, on the grounds that the
'experimental' intrinsic doesn't currently have much code generation
benefit, and my replacements encapsulate the sign- or zero-extension
so that you don't expose the illegal MVE vector type (`<4 x i64>`) in
IR.
The machine instructions come in two versions: with and without an
input accumulator. My new IR intrinsics, like the 'experimental' one,
don't take an accumulator parameter: we represent that by just adding
on the input value using an ordinary i32 or i64 add. So if you write
the `vaddvaq` C-language intrinsic with an input accumulator of zero,
it can be optimised to VADDV, and conversely, if you write something
like `x += vaddvq(y)` then that can be combined into VADDVA.
Most of this is achieved in isel lowering, by converting these IR
intrinsics into the existing `ARMISD::VADDV` family of custom SDNode
types. For the difficult case (64-bit accumulators), isel lowering
already implements the optimization of folding an addition into a
VADDLV to make a VADDLVA; so once we've made a VADDLV, our job is
already done, except that I had to introduce a parallel set of ARMISD
nodes for the //predicated// forms of VADDLV.
For the simpler VADDV, we handle the predicated form by just leaving
the IR intrinsic alone and matching it in an ordinary dag pattern.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, danielkiss, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76491
Summary:
I've implemented these as target-specific IR intrinsics, because
they're not //quite// enough like @llvm.experimental.vector.reduce.min
(which doesn't take the extra scalar parameter). Also this keeps the
predicated and unpredicated versions looking similar, and the
floating-point minnm/maxnm versions fold into the same schema.
We had a couple of min/max reductions already implemented, from the
initial pathfinding exercise in D67158. Those were done by having
separate IR intrinsic names for the signed and unsigned integer
versions; as part of this commit, I've changed them to use a flag
parameter indicating signedness, which is how we ended up deciding
that the rest of the MVE intrinsics family ought to work. So now
hopefully the ewhole lot is consistent.
In the new llc test, the output code from the `v8f16` test functions
looks quite unpleasant, but most of it is PCS lowering (you can't pass
a `half` directly in or out of a function). In other circumstances,
where you do something else with your `half` in the same function, it
doesn't look nearly as nasty.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: MarkMurrayARM
Subscribers: kristof.beyls, hiraditya, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76490
Summary:
This patch implements the following intrinsics:
uint8x16_t __arm_vcx1q_u8 (int coproc, uint32_t imm);
T __arm_vcx1qa(int coproc, T acc, uint32_t imm);
T __arm_vcx2q(int coproc, T n, uint32_t imm);
uint8x16_t __arm_vcx2q_u8(int coproc, T n, uint32_t imm);
T __arm_vcx2qa(int coproc, T acc, U n, uint32_t imm);
T __arm_vcx3q(int coproc, T n, U m, uint32_t imm);
uint8x16_t __arm_vcx3q_u8(int coproc, T n, U m, uint32_t imm);
T __arm_vcx3qa(int coproc, T acc, U n, V m, uint32_t imm);
Most of them are polymorphic. Furthermore, some intrinsics are
polymorphic by 2 or 3 parameter types, such polymorphism is not
supported by the existing MVE/CDE tablegen backends, also we don't
really want to have a combinatorial explosion caused by 1000 different
combinations of 3 vector types. Because of this some intrinsics are
implemented as macros involving a cast of the polymorphic arguments to
uint8x16_t.
The IR intrinsics are even more restricted in terms of types: all MVE
vectors are cast to v16i8.
Reviewers: simon_tatham, MarkMurrayARM, dmgreen, ostannard
Reviewed By: MarkMurrayARM
Subscribers: kristof.beyls, hiraditya, danielkiss, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76299
Summary:
This change implements ACLE CDE intrinsics that translate to
instructions working with general-purpose registers.
The specification is available at
https://static.docs.arm.com/101028/0010/ACLE_2019Q4_release-0010.pdf
Each ACLE intrinsic gets a corresponding LLVM IR intrinsic (because
they have distinct function prototypes). Dual-register operands are
represented as pairs of i32 values. Because of this the instruction
selection for these intrinsics cannot be represented as TableGen
patterns and requires custom C++ code.
Reviewers: simon_tatham, MarkMurrayARM, dmgreen, ostannard
Reviewed By: MarkMurrayARM
Subscribers: kristof.beyls, hiraditya, danielkiss, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76296
The MVE VDUP instruction take a GPR and splats into every lane of a
vector register. Unlike NEON we do not have a VDUPLANE equivalent
instruction, doing the same splat from a fp register. Previously a VDUP
to a v4f32/v8f16 would be represented as a (v4f32 VDUP f32), which
would mean the instruction pattern needs to add a COPY_TO_REGCLASS to
the GPR.
Instead this now converts that earlier during an ISel DAG combine,
converting (VDUP x) to (VDUP (bitcast x)). This can allow instruction
selection to tell that the input needs to be an i32, which in one of the
testcases allows it to use ldr (or specifically ldm) over (vldr;vmov).
Whilst being simple enough for floats, as the types sizes are the same,
these is no BITCAST equivalent for getting a half into a i32. This uses
a VMOVrh ARMISD node, which doesn't know the same tricks yet.
Differential Revision: https://reviews.llvm.org/D76292
Summary:
This is another set of instructions too complicated to be sensibly
expressed in IR by anything short of a target-specific intrinsic.
Given input vectors a,b, the instruction generates intermediate values
2*(a[0]*b[0]+a[1]+b[1]), 2*(a[2]*b[2]+a[3]+b[3]), etc; takes the high
half of each double-width values, and overwrites half the lanes in the
output vector c, which you therefore have to provide the input value
of. Optionally you can swap the elements of b so that the are things
like a[0]*b[1]+a[1]*b[0]; optionally you can round to nearest when
taking the high half; and optionally you can take the difference
rather than sum of the two products. Finally, saturation is applied
when converting back to a single-width vector lane.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: miyuki
Subscribers: kristof.beyls, hiraditya, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76359
When optimising for code size at the expense of performance, it is often
worth saving and restoring some of r0-r3, if IPRA will be able to take
advantage of them. This doesn't cost any extra code size if we already
have a PUSH/POP pair, and increases the number of available registers
across any calls to the function.
We already have an optimisation which tries fold the subtract/add of the
SP into the PUSH/POP by using extra registers, which somewhat conflicts
with this. I've made the new optimisation less aggressive in cases where
the existing one is likely to trigger, which gives better results than
either of these optimisations by themselves.
Differential revision: https://reviews.llvm.org/D69936
Summary:
These are complicated integer multiply+add instructions with extra
saturation, taking the high half of a double-width product, and
optional rounding. There's no sensible way to represent that in
standard IR, so I've converted the clang builtins directly to
target-specific intrinsics.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: miyuki
Subscribers: kristof.beyls, hiraditya, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76123
Summary:
These instructions compute multiply+add in integers, with one of the
operands being a splat of a scalar. (VMLA and VMLAS differ in whether
the splat operand is a multiplier or the addend.)
I've represented these in IR using existing standard IR operations for
the unpredicated forms. The predicated forms are done with target-
specific intrinsics, as usual.
When operating on n-bit vector lanes, only the bottom n bits of the
i32 scalar operand are used. So we have to tell that to isel lowering,
to allow it to remove a pointless sign- or zero-extension instruction
on that input register. That's done in `PerformIntrinsicCombine`, but
first I had to enable `PerformIntrinsicCombine` for MVE targets
(previously all the intrinsics it handled were for NEON), and make it
a method of `ARMTargetLowering` so that it can get at
`SimplifyDemandedBits`.
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, danielkiss, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76122
ISD::ROTL/ROTR rotation values are guaranteed to act as a modulo amount, so for power-of-2 bitwidths we only need the lowest bits.
Differential Revision: https://reviews.llvm.org/D76201
The ASRL/LSRL long shifts are generated from 64bit shifts. Once we have
them, it might turn out that enough of the 64bit result was not required
that we can use a smaller shift to perform the same result. As the
smaller shift can in general be folded in more way, such as into add
instructions in one of the test cases here, we can use the demand bit
analysis to prefer the smaller shifts where we can.
Differential Revision: https://reviews.llvm.org/D75371
This changes the way that asrl and lsrl intrinsics are lowered, going
via a the ISEL ASRL and LSLL nodes instead of straight to machine nodes.
On top of that, it adds some constant folds for long shifts, in case it
turns out that the shift amount was either constant or 0.
Differential Revision: https://reviews.llvm.org/D75553
Summary:
This adds the ACLE intrinsic family for the VFMA and VFMS
instructions, which perform fused multiply-add on vectors of floats.
I've represented the unpredicated versions in IR using the cross-
platform `@llvm.fma` IR intrinsic. We already had isel rules to
convert one of those into a vector VFMA in the simplest possible way;
but we didn't have rules to detect a negated argument and turn it into
VFMS, or rules to detect a splat argument and turn it into one of the
two vector/scalar forms of the instruction. Now we have all of those.
The predicated form uses a target-specific intrinsic as usual, but
I've stuck to just one, for a predicated FMA. The subtraction and
splat versions are code-generated by passing an fneg or a splat as one
of its operands, the same way as the unpredicated version.
In arm_mve_defs.h, I've had to introduce a tiny extra piece of
infrastructure: a record `id` for use in codegen dags which implements
the identity function. (Just because you can't declare a Tablegen
value of type dag which is //only// a `$varname`: you have to wrap it
in something. Now I can write `(id $varname)` to get the same effect.)
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, danielkiss, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D75998
Tim Northover