Current tail duplication integrated in bb layout is designed to increase the fallthrough from a BB's predecessor to its successor, but we have observed cases that duplication doesn't increase fallthrough, or it brings too much size overhead.
To overcome these two issues in function canTailDuplicateUnplacedPreds I add two checks:
make sure there is at least one duplication in current work set.
the number of duplication should not exceed the number of successors.
The modification in hasBetterLayoutPredecessor fixes a bug that potential predecessor must be at the bottom of a chain.
Differential Revision: https://reviews.llvm.org/D64376
The VCMP instructions in MVE can accept a register or ZR, but only as
the right hand operator. Most of the time this will already be correct
because the icmp will have been canonicalised that way already. There
are some cases in the lowering of float conditions that this will not
apply to though. This code should fix up those cases.
Differential Revision: https://reviews.llvm.org/D70822
Summary:
This commit adds the `vpselq` intrinsics which take an MVE predicate
word and select lanes from two vectors; the `vctp` intrinsics which
create a tail predicate word suitable for processing the first m
elements of a vector (e.g. in the last iteration of a loop); and
`vpnot`, which simply complements a predicate word and is just
syntactic sugar for the `~` operator.
The `vctp` ACLE intrinsics are lowered to the IR intrinsics we've
already added (and which D70592 just reorganized). I've filled in the
missing isel rule for VCTP64, and added another set of rules to
generate the predicated forms.
I needed one small tweak in MveEmitter to allow the `unpromoted` type
modifier to apply to predicates as well as integers, so that `vpnot`
doesn't pointlessly convert its input integer to an `<n x i1>` before
complementing it.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D70485
Summary:
D65884 added a set of Arm IR intrinsics for the MVE VCTP instruction,
to use in tail predication. But the 64-bit one doesn't work properly:
its predicate type is `<2 x i1>` / `v2i1`, which isn't a legal MVE
type (due to not having a full set of instructions that manipulate it
usefully). The test of `vctp64` in `basic-tail-pred.ll` goes through
`opt` fine, as the test expects, but if you then feed it to `llc` it
causes a type legality failure at isel time.
The usual workaround we've been using in the rest of the MVE
intrinsics family is to bodge `v2i1` into `v4i1`. So I've adjusted the
`vctp64` IR intrinsic to do that, and completely removed the code (and
test) that uses that intrinsic for 64-bit tail predication. That will
allow me to add isel rules (upcoming in D70485) that actually generate
the VCTP64 instruction.
Also renamed all four of these IR intrinsics so that they have `mve`
in the name, since its absence was confusing.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: MarkMurrayARM
Subscribers: samparker, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70592
Summary:
If a user writing C code using the ACLE MVE intrinsics generates a
predicate and then complements it, then the resulting IR will use the
`pred_v2i` IR intrinsic to turn some `<n x i1>` vector into a 16-bit
integer; complement that integer; and convert back. This will generate
machine code that moves the predicate out of the `P0` register,
complements it in an integer GPR, and moves it back in again.
This InstCombine rule replaces `i2v(~v2i(x))` with a direct complement
of the original predicate vector, which we can already instruction-
select as the VPNOT instruction which complements P0 in place.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70484
These instructions do not work quite like I expected them to. They
perform the addition and then shift in a higher precision integer, so do
not match up with the patterns that we added.
For example with s8s, adding 100 and 100 should wrap leaving the shift
to work on a negative number. VHADD will instead do the arithmetic in
higher precision, giving 100 overall. The vhadd gives a "better" result,
but not one that matches up with the input.
I am just removing the patterns here. We might be able to re-add them in
the future by checking for wrap flags or changing bitwidths. But for the
moment just remove them to remove the problem cases.
MVE has a basic symmetry between it's normal loads/store operations and
the masked variants. This means that masked loads and stores can use
pre-inc and post-inc addressing modes, just like the standard loads and
stores already do.
To enable that, this patch adds all the relevant infrastructure for
treating masked loads/stores addressing modes in the same way as normal
loads/stores.
This involves:
- Adding an AddressingMode to MaskedLoadStoreSDNode, along with an extra
Offset operand that is added after the PtrBase.
- Extending the IndexedModeActions from 8bits to 16bits to store the
legality of masked operations as well as normal ones. This array is
fairly small, so doubling the size still won't make it very large.
Offset masked loads can then be controlled with
setIndexedMaskedLoadAction, similar to standard loads.
- The same methods that combine to indexed loads, such as
CombineToPostIndexedLoadStore, are adjusted to handle masked loads in
the same way.
- The ARM backend is then adjusted to make use of these indexed masked
loads/stores.
- The X86 backend is adjusted to hopefully be no functional changes.
Differential Revision: https://reviews.llvm.org/D70176
Add some more helper functions to ReachingDefs to query the uses of
a given MachineInstr and also to query whether two MachineInstrs use
the same def of a register.
For Arm, while tail-predicating, these helpers are used in the
low-overhead loops to remove the dead code that calculates the number
of loop iterations.
Differential Revision: https://reviews.llvm.org/D70240
Add several new methods to ReachingDefAnalysis:
- getReachingMIDef, instead of returning an integer, return the
MachineInstr that produces the def.
- getInstFromId, return a MachineInstr for which the given integer
corresponds to.
- hasSameReachingDef, return whether two MachineInstr use the same
def of a register.
- isRegUsedAfter, return whether a register is used after a given
MachineInstr.
These methods have been used in ARMLowOverhead to replace searching
for uses/defs.
Differential Revision: https://reviews.llvm.org/D70009
When inserting a non-decrementing LE, the basic block was being
resized to take into consideration that a tCMP and tBcc had been
combined into one T1 instruction. This is not true in the LE case
where we generate a CBN?Z and an LE.
Differential Revision: https://reviews.llvm.org/D70536
Adds a pattern to ARMInstrMVE.td to use a VQNEG
instruction if an equivalent multi-instruction
construct is found.
Differential Revision: https://reviews.llvm.org/D70491
Adds a pattern to ARMInstrMVE.td to use a VQABS
instruction if an equivalent multi-instruction
construct is found.
Differential revision: https://reviews.llvm.org/D70181
Now that we have the intrinsics, we can add VLD2/4 and VST2/4 lowering
for MVE. This works the same way as Neon, recognising the load/shuffles
combination and converting them into intrinsics in a pre-isel pass,
which just calls getMaxSupportedInterleaveFactor, lowerInterleavedLoad
and lowerInterleavedStore.
The main difference to Neon is that we do not have a VLD3 instruction.
Otherwise most of the code works very similarly, with just some minor
differences in the form of the intrinsics to work around. VLD3 is
disabled by making isLegalInterleavedAccessType return false for those
cases.
We may need some other future adjustments, such as VLD4 take up half the
available registers so should maybe cost more. This patch should get the
basics in though.
Differential Revision: https://reviews.llvm.org/D69392
This fills in the small family of MVE intrinsics that have nothing to
do with vectors: they implement bit-shift operations on 32- or 64-bit
values held in one or two general-purpose registers. Most of these
shift operations saturate if shifting left, and round to nearest if
shifting right, although LSLL and ASRL behave like ordinary shifts.
When these instructions take a variable shift count in a register,
they pay attention to its sign, so that (for example) LSLL or UQRSHLL
will shift left if given a positive number but right if given a
negative one. That makes even LSLL and ASRL different enough from
standard LLVM IR shift semantics that I couldn't see any better
alternative than to simply model the whole family as a set of
MVE-specific IR intrinsics.
(The //immediate// forms of LSLL and ASRL, on the other hand, do
behave exactly like a standard IR shift of a 64-bit value. In fact,
those forms don't have ACLE intrinsics defined at all, because you can
just write an ordinary C shift operation if you want one of those.)
The 64-bit shifts have to be instruction-selected in C++, because they
deliver two output values. But the 32-bit ones are simple enough that
I could write a DAG isel pattern directly into each Instruction
record.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D70319
Remove the restriction, from the mve tail predication pass, that the
all masked vectors instructions need to be 128-bits. This allows us
to supported extending loads and truncating stores.
Differential Revision: https://reviews.llvm.org/D69946
This patch modifies ARMLowOverheadLoops to convert a predicated
vector low-overhead loop into a tail-predicatd one. This is currently
a very basic conversion, with the following restrictions:
- Operates only on single block loops.
- The loop can only contain a single vctp instruction.
- No other instructions can write to the vpr.
- We only allow a subset of the mve instructions in the loop.
TODO: Pass the number of elements, not the number of iterations to
dlstp/wlstp.
Differential Revision: https://reviews.llvm.org/D69945
If you're writing C code using the ACLE MVE intrinsics that passes the
result of a vcmp as input to a predicated intrinsic, e.g.
mve_pred16_t pred = vcmpeqq(v1, v2);
v_out = vaddq_m(v_inactive, v3, v4, pred);
then clang's codegen for the compare intrinsic will create calls to
`@llvm.arm.mve.pred.v2i` to convert the output of `icmp` into an
`mve_pred16_t` integer representation, and then the next intrinsic
will call `@llvm.arm.mve.pred.i2v` to convert it straight back again.
This will be visible in the generated code as a `vmrs`/`vmsr` pair
that move the predicate value pointlessly out of `p0` and back into it again.
To prevent that, I've added InstCombine rules to remove round trips of
the form `v2i(i2v(x))` and `i2v(v2i(x))`. Also I've taught InstCombine
about the known and demanded bits of those intrinsics. As a result,
you now get just the generated code you wanted:
vpt.u16 eq, q1, q2
vaddt.u16 q0, q3, q4
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70313
Summary:
As well as vector/vector compare instructions, MVE also has a family
of comparisons taking a vector and a scalar, which compare every lane
of the vector against the same value. We generate those at isel time
using isel patterns that match `(ARMvcmp vector, (ARMvdup scalar))`.
This commit adds corresponding patterns for the operand-reversed form
`(ARMvcmp (ARMvdup scalar), vector)`, with condition codes swapped as
necessary. That way, we can still generate the vector/scalar compare
instruction if the IR happens to have been rearranged to put the
operands the other way round, which can happen in some optimization
phases. Previously, a vcmp the other way round was handled by emitting
a `vdup` instruction to //explicitly// replicate the scalar input into
a vector, and then doing a vector/vector comparison.
I haven't added a new test, because it turned out that several
existing tests were already exhibiting that failure mode. So just
updating the expected output in the existing MVE codegen tests
demonstrates what's been improved.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Reviewed By: dmgreen
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70296
This patch adds the ACLE intrinsics for all the MVE load and store
instructions not already handled by D69791. These ones don't need new
IR intrinsics, because they can be implemented in terms of standard
LLVM IR constructions.
Some of the load and store instructions access less than 128 bits of
memory, sign/zero extending each value to a wider vector lane on load
or truncating it on store. These are represented in IR by a load of a
shorter vector followed by a zext/sext, and conversely, a trunc
followed by a short store. Existing ISel patterns already recognize
those combinations and turn them into the right MVE instructions.
The predicated forms of all these instructions are represented in the
same way, except that the ordinary load/store operation is replaced
with the existing intrinsics @llvm.masked.{load,store}. These are
currently only code-generated as predicated MVE load/store
instructions if you give LLVM the `-enable-arm-maskedldst` option; so
I've done that in the LLVM codegen test. When we make that the
default, that option can be removed.
In the Tablegen backend, I've had to add a handful of extra support
features:
* We need to be able to make clang::Address objects out of a
pointer and an alignment (previously we only needed these when the
user passed us an existing one).
* We can now specify vector types that aren't 128 bits wide (for use
in those intermediate values in IR), the parametrized type system
can make one starting from two existing vector types (using the lane
count of one and the element type of the other).
* I've added support for code generation of pointer casts, and for
specifying LLVM types as operands to IRBuilder operations (for zext
and sext, though I think they'll come in useful again).
* Now not all IR construction operations need to be specified as
Builder.CreateFoo; some don't involve a Builder at all, and one
passes it as a parameter to a tiny static helper function in
CGBuiltin.cpp.
Reviewers: ostannard, MarkMurrayARM, dmgreen
Subscribers: kristof.beyls, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D70088
MVE includes instructions that extract an 8- or 16-bit lane from a
vector and sign-extend it into the output 32-bit GPR. `ARMInstrMVE.td`
already included isel patterns to select those instructions in
response to the `ARMISD::VGETLANEs` selection-DAG node type. But
`ARMISD::VGETLANEs` was never actually generated, because the code
that creates it was conditioned on NEON only.
It's an easy fix to enable the same code for integer MVE, and now IR
that sign-extends the result of an extractelement (whether explicitly
or as part of the function call ABI) will use `vmov.s8` instead of
`vmov.u8` followed by `sxtb`.
Reviewers: SjoerdMeijer, dmgreen, ostannard
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70132
This patch adds two new families of intrinsics, both of which are
memory accesses taking a vector of locations to load from / store to.
The vldrq_gather_base / vstrq_scatter_base intrinsics take a vector of
base addresses, and an immediate offset to be added consistently to
each one. vldrq_gather_offset / vstrq_scatter_offset take a scalar
base address, and a vector of offsets to add to it. The
'shifted_offset' variants also multiply each offset by the element
size type, so that the vector is effectively of array indices.
At the IR level, these operations are represented by a single set of
four IR intrinsics: {gather,scatter} × {base,offset}. The other
details (signed/unsigned, shift, and memory element size as opposed to
vector element size) are all specified by IR intrinsic polymorphism
and immediate operands, because that made the selection job easier
than making a huge family of similarly named intrinsics.
I considered using the standard IR representations such as
llvm.masked.gather, but they're not a good fit. In order to use
llvm.masked.gather to represent a gather_offset load with element size
smaller than a pointer, you'd have to expand the <8 x i16> vector of
offsets into an <8 x i16*> vector of pointers, which would be split up
during legalization, so you'd spend most of your time undoing the mess
it had made. Also, ISel support for llvm.masked.gather would be easy
enough in a trivial way (you can expand it into a gather-base load
with a zero immediate offset), but instruction-selecting lots of
fiddly idioms back into all the _other_ MVE load instructions would be
much more work. So I think dedicated IR intrinsics are the more
sensible approach, at least for the moment.
On the clang tablegen side, I've added two new features to the
Tablegen source accepted by MveEmitter: a 'CopyKind' type node for
defining a type that varies with the parameter type (it lets you ask
for an unsigned integer type of the same width as the parameter), and
an 'unsignedflag' value node for passing an immediate IR operand which
is 0 for a signed integer type or 1 for an unsigned one. That lets me
write each kind of intrinsic just once and get all its subtypes and
immediate arguments generated automatically.
Also I've tweaked the handling of pointer-typed values in the code
generation part of MveEmitter: they're generated as Address rather
than Value (i.e. including an alignment) so that they can be given to
the ordinary IR load and store operations, but I'd omitted the code to
convert them back to Value when they're going to be used as an
argument to an IR intrinsic.
On the MC side, I've enhanced MVEVectorVTInfo so that it can tell you
not only the full assembly-language suffix for a given vector type
(like 's32' or 'u16') but also the numeric-only one used by store
instructions (just '32' or '16').
Reviewers: dmgreen
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D69791
The Arm backend will usually return false for isFMAFasterThanFMulAndFAdd,
where both the fused VFMA.f32 and a non-fused VMLA.f32 are usually
available for scalar code. For MVE we don't have the non-fused version
though. It makes more sense for isFMAFasterThanFMulAndFAdd to return
true, allowing us to simplify some of the existing ISel patterns.
The tests here are that non of the existing tests failed, and so we are
still selecting VFMA and VFMS. The one test that changed shows we can
now select from fast math flags, as opposed to just relying on the
isFMADLegalForFAddFSub option.
Differential Revision: https://reviews.llvm.org/D69115
The MVE VADC instruction reads and writes the carry bit at bit 29 of
the FPSCR register. The corresponding ACLE intrinsic is specified to
work with an integer in which the carry bit is stored at bit 0. So if
a user writes a code sequence in C that passes the carry from one VADC
to the next, like this,
s0 = vadcq_u32(a0, b0, &carry);
s1 = vadcq_u32(a1, b1, &carry);
then clang will generate IR for each of those operations that shifts
the carry bit up into bit 29 before the VADC, and after it, shifts it
back down and masks off all but the low bit. But in this situation
what you really wanted was two consecutive VADC instructions, so that
the second one directly reads the value left in FPSCR by the first,
without wasting several instructions on pointlessly clearing the other
flag bits in between.
This commit explains to InstCombine that the other bits of the flags
operand don't matter, and adds a test that demonstrates that all the
code between the two VADC instructions can be optimized away as a
result.
Reviewers: dmgreen, miyuki, ostannard
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67162
The VST2 and VST4 instructions take two or four vector registers as
input, and store part of each register to memory in an interleaved
pattern. They come in variants indicating which part of each register
they store (VST20 and VST21; VST40 to VST43 inclusive); the intention
is that issuing each of those variants in turn has the combined effect
of loading or storing the whole set of registers to a memory block of
equal size. The corresponding VLD2 and VLD4 instructions load from
memory in the same interleaved format: each one overwrites only part
of its output register set, and again, the idea is that if you use
VLD4{0,1,2,3} or VLD2{0,1} together, you end up having written to the
whole of each register.
I've implemented the stores and loads quite differently. The loads
were easiest to implement as a single intrinsic that expands to all
four VLD4x instructions or both VLD2x, delivering four complete output
registers. (Implementing each individual load as a separate
instruction taking four input registers to partially overwrite is
possible in theory, but pointless, and when I tried it, I found it
would need extra work to get the register allocation not to be
horrible.) Since that intrinsic delivers multiple outputs, it has to
be instruction-selected in custom C++.
But the store instructions are easier to model individually, because
they don't overwrite any register at all and you can write a DAG Isel
pattern in Tablegen for each one.
Hence, my new intrinsic `int_arm_mve_vld4q` expands to four load
instructions, delivers four full output vectors, and is handled by C++
code, whereas `int_arm_mve_vst4q` expands to just one store
instruction, takes four input vectors and a constant indicating which
lanes to store, and is handled entirely in Tablegen. (And similarly
for vld2q/vst2q.) This is asymmetric, but it was the easiest way to do
each one.
Reviewers: dmgreen, miyuki, ostannard
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68700
This adds some initial example IR intrinsics for MVE instructions that
deliver multiple output values, and hence, have to be instruction-
selected by custom C++ code instead of Tablegen patterns.
I've added the writeback gather load instructions (taking a vector of
base addresses and a single common offset, returning a vector of
loaded values and an updated vector of base addresses); one example
from the long shift family (taking and returning a 64-bit value in two
GPRs); and the VADC instruction (which propagates a carry bit from
each vector-lane addition to the next, taking an input carry flag in
FPSCR and outputting the final one in FPSCR as well).
To support the VPT-predicated forms of these instructions, I've
written some helper functions to add the cluster of MVE predicate
operands to the end of a MachineInstr. `AddMVEPredicateToOps` is used
when the instruction actually is predicated (so it takes a predicate
mask argument), and `AddEmptyMVEPredicateToOps` is for when the
instruction is unpredicated (so it fills in $noreg for the mask). Each
one comes in a form suitable for `vpred_n`, and one for `vpred_r`
which takes the extra 'inactive' parameter.
For VADC, the representation of the carry flag in the IR intrinsic is
a word intended to be moved directly to and from `FPSCR_nzcvqc`, i.e.
with the carry flag in bit 29 of the word. (The user-facing ACLE
intrinsic will want it to be in bit 0, but I'll do that on the clang
side.)
Reviewers: dmgreen, miyuki, ostannard
Subscribers: kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68699
This commit, together with the next few, will add a representative
sample of the kind of IR intrinsics that we'll need in order to
implement the user-facing ACLE intrinsics for MVE. Supporting all of
them will take more work; the intention of this initial series of
commits is to implement an intrinsic or two from lots of different
categories, as examples and proofs of concept.
This initial commit introduces a small number of IR intrinsics for
instructions simple enough that they can use Tablegen ISel patterns:
the predicated versions of the VADD and VSUB instructions (both
integer and FP), VMIN and VMAX, and the float->half VCVT instruction
(predicated and unpredicated).
When using VPT-predicated instructions in automatic code generation,
it will be convenient to specify the predicate value as a vector of
the appropriate number of i1. To make it easy to specify all sizes of
an instruction in one go and give each one the matching predicate
vector type, I've added a system of Tablegen informational records
describing MVE's vector types: each one gives the underlying LLVM IR
ValueType (which may not be the same if the MVE vector is of
explicitly signed or unsigned integers) and an appropriate vNi1 to use
as the predicate vector.
(Also, those info records include the usual encoding for the types, so
that as we add associations between each instruction encoding and one
of the new `MVEVectorVTInfo` records, we can remove some of the
existing template parameters and replace them with references to the
vector type info's fields.)
The user-facing ACLE intrinsics will receive a predicate mask as a
16-bit integer, so I've also provided a pair of intrinsics i2v and
v2i, to convert between an integer and a vector of i1 by just changing
the register class.
Reviewers: dmgreen, miyuki, ostannard
Subscribers: javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67158
Allow us to generate truncating masked store which take v4i32 and
v8i16 vectors and can store to v4i8, v4i16 and v8i8 and memory.
Removed support for unaligned masked stores.
Differential Revision: https://reviews.llvm.org/D68461
llvm-svn: 375108
Add generic DAG combine for extending masked loads.
Allow us to generate sext/zext masked loads which can access v4i8,
v8i8 and v4i16 memory to produce v4i32, v8i16 and v4i32 respectively.
Differential Revision: https://reviews.llvm.org/D68337
llvm-svn: 375085
The adds both VMOVNt and VMOVNb instruction selection from the appropriate
shuffles. We detect shuffle masks of the form:
0, N, 2, N+2, 4, N+4, ...
or
0, N+1, 2, N+3, 4, N+5, ...
ISel will also try the opposite patterns, with inputs reversed. These are
selected to VMOVNt and VMOVNb respectively.
Differential Revision: https://reviews.llvm.org/D68283
llvm-svn: 374781
This selects MVE VQADD from the vector llvm.sadd.sat or llvm.uadd.sat
intrinsics.
Differential Revision: https://reviews.llvm.org/D68566
llvm-svn: 374336
Based on the discussion in
http://lists.llvm.org/pipermail/llvm-dev/2019-October/135574.html, the
conclusion was reached that the ARM backend should produce vcmp instead
of vcmpe instructions by default, i.e. not be producing an Invalid
Operation exception when either arguments in a floating point compare
are quiet NaNs.
In the future, after constrained floating point intrinsics for floating
point compare have been introduced, vcmpe instructions probably should
be produced for those intrinsics - depending on the exact semantics
they'll be defined to have.
This patch logically consists of the following parts:
- Revert http://llvm.org/viewvc/llvm-project?rev=294945&view=rev and
http://llvm.org/viewvc/llvm-project?rev=294968&view=rev, which
implemented fine-tuning for when to produce vcmpe (i.e. not do it for
equality comparisons). The complexity introduced by those patches
isn't needed anymore if we just always produce vcmp instead. Maybe
these patches need to be reintroduced again once support is needed to
map potential LLVM-IR constrained floating point compare intrinsics to
the ARM instruction set.
- Simply select vcmp, instead of vcmpe, see simple changes in
lib/Target/ARM/ARMInstrVFP.td
- Adapt lots of tests that tested for vcmpe (instead of vcmp). For all
of these test, the intent of what is tested for isn't related to
whether the vcmp should produce an Invalid Operation exception or not.
Fixes PR43374.
Differential Revision: https://reviews.llvm.org/D68463
llvm-svn: 374025
Darwin platforms need the frame register to always point at a valid record even
if it's not updated in a leaf function. Backtraces are more important than one
extra GPR.
llvm-svn: 373738
Identity shuffles, of the form (0, 1, 2, 3, ...) are perfectly OK under MVE
(they essentially just become bitcasts). We were not catching that in the
existing set of what we considered legal though. On NEON, they would be covered
by vext's, but that is not generally available in MVE.
This uses ShuffleVectorInst::isIdentityMask which is a little odd to use here
but does what we want and prevents us from just rewriting what is the same
function.
Differential Revision: https://reviews.llvm.org/D68241
llvm-svn: 373446
The VCTP instruction will calculate the predicate masked based upon
the number of elements that need to be processed. I had inserted the
sub before the vctp intrinsic and supplied it as the operand, but
this is incorrect as the phi should directly feed the vctp. The sub
is calculating the value for the next iteration.
Differential Revision: https://reviews.llvm.org/D67921
llvm-svn: 373188
During legalisation we can end up with some pretty strange nodes, like shifts
of 0. We need to make sure we don't try to make long shifts of these, ending up
with invalid assembly instructions. A long shift with a zero immediate actually
encodes a shift by 32.
Differential Revision: https://reviews.llvm.org/D67664
llvm-svn: 372839
Similar to rL372717, we can force the splitting of extends of vector loads in
MVE, in order to use the better widening loads as opposed to going through
expensive extends. This adds a combine to early-on detect extends of loads and
split the load in two, from where normal legalisation will kick in and we get a
series of widening loads.
Differential Revision: https://reviews.llvm.org/D67909
llvm-svn: 372721
MVE does not have a simple sign extend instruction that can move elements
across lanes. We currently often end up moving each lane into and out of a GPR,
in order to get elements into the correct places. When we have a store of a
trunc (or a extend of a load), we can instead just split the store/load in two,
using the narrowing/widening load/store instructions from each half of the
vector.
This does that for stores. It happens very early in a store combine, so as to
easily detect the truncates. (It would be possible to do this later, but that
would involve looking through a buildvector of extract elements. Not impossible
but this way seemed simpler).
By enabling store combines we also get a vmovdrr combine for free, helping some
other tests.
Differential Revision: https://reviews.llvm.org/D67828
llvm-svn: 372717
Remove any predicate that we replace with a vctp intrinsic, and try
to remove their operands too. Also look into the exit block to see if
there's any duplicates of the predicates that we've replaced and
clone the vctp to be used there instead.
Differential Revision: https://reviews.llvm.org/D67709
llvm-svn: 372567
Check whether there are any uses or defs between the LoopDec and
LoopEnd. If there's not, then we can use a subs to set the cpsr and
skip generating a cmp.
Differential Revision: https://reviews.llvm.org/D67801
llvm-svn: 372560
We needn't BFI each lane individually into a predicate register when each lane
in the same. A simple sign extend and a vmsr will do.
Differential Revision: https://reviews.llvm.org/D67653
llvm-svn: 372313
We were previously using the SelectT2AddrModeImm7 for both normal and narrowing
MVE loads/stores. As the narrowing instructions do not accept sp as a register,
it makes little sense to optimise a FrameIndex into the load, only to have to
recover that later on. This adds a SelectTAddrModeImm7 which does not do that
folding, and uses it for narrowing load/store patterns.
Differential Revision: https://reviews.llvm.org/D67489
llvm-svn: 372134
Similar to D67327, but this time for the FP16 VLDR and VSTR instructions that
use the AddrMode5FP16 addressing mode. We need to reserve an emergency spill
slot for instructions that will be out of range to use sp directly.
AddrMode5FP16 is 8 bits with a scale of 2.
Differential Revision: https://reviews.llvm.org/D67483
llvm-svn: 372132
Remove setPreservesCFG from ARMConstantIslandPass and add a couple
of -verify-machine-dom-info instances into the existing codegen
tests.
llvm-svn: 372126
MVE loads and stores have a 7 bit immediate range, scaled by the length of the type. This needs to be taught to the stack estimation code to ensure that an emergency spill slot is reserved in case we run out of registers when materialising stack indices.
Also the narrowing loads/stores can be created with frame indices even though they do not accept SP as a register. We need in those cases to make sure we have an emergency register to use as the frame base, as SP can never be used.
Differential Revision: https://reviews.llvm.org/D67327
llvm-svn: 372114
Converting the *LoopStart pseudo instructions into DLS/WLS results in
LR being defined. These instructions were inserted on the assumption
that LR would already contain the loop counter because a mov is
introduced during ISel as the the consumers in the loop can only use
LR. That assumption proved wrong!
So perform a safety check, finding an appropriate place to insert the
DLS/WLS instructions or revert if this isn't possible.
Differential Revision: https://reviews.llvm.org/D67539
llvm-svn: 372111
The low-overhead branch extension provides a loop-end 'LE' instruction
that performs no decrement nor compare, it just jumps backwards. This
patch modifies the constant islands pass to try to insert LE
instructions in place of a Thumb2 conditional branch, instead of
shrinking it. This only happens if a cmp can be converted to a cbn/z
and used to exit the loop.
Differential Revision: https://reviews.llvm.org/D67404
llvm-svn: 372085
The adds some very basic folding of PREDICATE_CASTS, removing cases when they
are chained together. These would already be removed eventually, as these are
lowered to copies. This just allows it to happen earlier, which can help other
simplifications.
Differential Revision: https://reviews.llvm.org/D67591
llvm-svn: 372012
Lower CTTZ on MVE using VBRSR and VCLS which will reverse the bits and
count the leading zeros, equivalent to a count trailing zeros (CTTZ).
llvm-svn: 372000
MVE has VPT instructions, which perform the duties of both a VCMP and a VPST in
a single instruction, performing the compare and starting the VPT block in one.
This teaches the MVEVPTBlockPass to fold them, searching back through the
basicblock for a valid VCMP and creating the VPT from its operands.
There are some changes to the VPT instructions to accommodate this, altering
the order of the operands to match the VCMP better, and changing P0 register
defs to be VPR defs, as is used in other places.
Differential Revision: https://reviews.llvm.org/D66577
llvm-svn: 371982
Masked loads and store fit naturally with MVE, the instructions being easily
predicated. This adds lowering for the simple cases of masked loads and stores.
It does not yet deal with widening/narrowing or pre/post inc, and so is
currently behind an option.
The llvm masked load intrinsic will accept a "passthru" value, dictating the
values used for the zero masked lanes. In MVE the instructions write 0 to the
zero predicated lanes, so we need to match a passthru that isn't 0 (or undef)
with a select instruction to pull in the correct data after the load.
Differential Revision: https://reviews.llvm.org/D67186
llvm-svn: 371932
This patch adds vecreduce_smax, vecredude_umax, vecreduce_smin, vecreduce_umin and selection for vmaxv and minv.
Differential Revision: https://reviews.llvm.org/D66413
llvm-svn: 371827
Summary:
This catches malformed mir files which specify alignment as log2 instead of pow2.
See https://reviews.llvm.org/D65945 for reference,
This is patch is part of a series to introduce an Alignment type.
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2019-July/133851.html
See this patch for the introduction of the type: https://reviews.llvm.org/D64790
Reviewers: courbet
Subscribers: MatzeB, qcolombet, dschuff, arsenm, sdardis, nemanjai, jvesely, nhaehnle, hiraditya, kbarton, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, MaskRay, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, PkmX, jocewei, jsji, Petar.Avramovic, asbirlea, s.egerton, pzheng, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67433
llvm-svn: 371608
These predicate vectors can usually be loaded and stored with a single
instruction, a VSTR_P0. However this instruction will store the entire P0
predicate, 16 bits, zeroextended to 32bits. Each lane of the the
v4i1/v8i1/v16i1 representing 4/2/1 bits.
As far as I understand, when llvm says "store this v4i1", it really does need
to store 4 bits (or 8, that being the size of a byte, with this bottom 4 as the
interesting bits). For example a bitcast from a v8i1 to a i8 is defined as a
store followed by a load, which is how the code is expanded.
So this instead lowers the v4i1/v8i1 load/store through some shuffles to get
the bits into the correct positions. This, as you might imagine, is not as
efficient as a single instruction. But I believe it is needed for correctness.
v16i1 equally should not load/store 32bits, only storing the 16bits of data.
Stack loads/stores are still using the VSTR_P0 (as can be seen by the test not
changing). This is fine as they are self-consistent, it is only "externally
observable loads/stores" (from our point of view) that need to be corrected.
Differential revision: https://reviews.llvm.org/D67085
llvm-svn: 371419
This patch sinks add/mul(shufflevector(insertelement())) into the basic block in which they are used so that they can then be selected together.
This is useful for various MVE instructions, such as vmla and others that take R registers.
Loop tests have been added to the vmla test file to make sure vmlas are generated in loops.
Differential revision: https://reviews.llvm.org/D66295
llvm-svn: 371218
The MVE and LOB extensions of Armv8.1m can be combined to enable
'tail predication' which removes the need for a scalar remainder
loop after vectorization. Lane predication is performed implicitly
via a system register. The effects of predication is described in
Section B5.6.3 of the Armv8.1-m Arch Reference Manual, the key points
being:
- For vector operations that perform reduction across the vector and
produce a scalar result, whether the value is accumulated or not.
- For non-load instructions, the predicate flags determine if the
destination register byte is updated with the new value or if the
previous value is preserved.
- For vector store instructions, whether the store occurs or not.
- For vector load instructions, whether the value that is loaded or
whether zeros are written to that element of the destination
register.
This patch implements a pass that takes a hardware loop, containing
masked vector instructions, and converts it something that resembles
an MVE tail predicated loop. Currently, if we had code generation,
we'd generate a loop in which the VCTP would generate the predicate
and VPST would then setup the value of VPR.PO. The loads and stores
would be placed in VPT blocks so this is not tail predication, but
normal VPT predication with the predicate based upon a element
counting induction variable. Further work needs to be done to finally
produce a true tail predicated loop.
Because only the loads and stores are predicated, in both the LLVM IR
and MIR level, we will restrict support to only lane-wise operations
(no horizontal reductions). We will perform a final check on MIR
during loop finalisation too.
Another restriction, specific to MVE, is that all the vector
instructions need operate on the same number of elements. This is
because predication is performed at the byte level and this is set
on entry to the loop, or by the VCTP instead.
Differential Revision: https://reviews.llvm.org/D65884
llvm-svn: 371179
This attempts to just fix the creation of VPT blocks, fixing up the iterating,
which instructions are considered in the bundle, and making sure that we do not
overrun the end of the block.
Differential Revision: https://reviews.llvm.org/D67219
llvm-svn: 371064
The code here seems to date back to r134705, when tablegen lowering was first
being added. I don't believe that we need to include CPSR implicit operands on
the MCInst. This now works more like other backends (like AArch64), where all
implicit registers are skipped.
This allows the AliasInst for CSEL's to match correctly, as can be seen in the
test changes.
Differential revision: https://reviews.llvm.org/D66703
llvm-svn: 370745
This moves ConstantMaterializationCost into ARMBaseInstrInfo so that it can
also be used in ISel Lowering, adding codesize values to the computed costs, to
be able to compare either approximate instruction counts or codesize costs.
It also adds a HasLowerConstantMaterializationCost, which compares the
ConstantMaterializationCost of two values, returning true if the first is
smaller either in instruction count/codesize, or falling back to the other in
the case that they are equal.
This is used in constant CSEL lowering to invert the predicate if the opposite
is easier to materialise.
Differential revision: https://reviews.llvm.org/D66701
llvm-svn: 370741
Arm 8.1-M adds a number of related CSEL instructions, including CSINC, CSNEG and CSINV. These choose between two values given the content in CPSR and a condition, performing an increment, negation or inverse of the false value.
This adds some selection for them, either from constant values or patterns. It does not include CSEL directly, which is currently not always making code better. It is still useful, but we will have to check more carefully where it should and shouldn't be used.
Code by Ranjeet Singh and Simon Tatham, with some modifications from me.
Differential revision: https://reviews.llvm.org/D66483
llvm-svn: 370739
We were using isShiftedInt<7, Shift>(RHSC) to detect the ranges of offsets to
fold into MVE loads/stores. The instructions actually take a 7 bit unsigned
integer which is either added or subtracted. So something more like
isShiftedUInt<7, Shift>(abs(RHSC)).
Instead I've changes this to use the isScaledConstantInRange method, same as in
SelectT2AddrModeImm7Offset used by pre/post inc, which seemed to already be
getting this correct.
Differential revision: https://reviews.llvm.org/D66997
llvm-svn: 370731
We should be using MQPR, and if we don't we can get COPYs and PHIs created for
QPR. These get folded into instructions, failing verification checks.
Differential revision: https://reviews.llvm.org/D66214
llvm-svn: 370676
The motivating case for this is a long way from here:
https://bugs.llvm.org/show_bug.cgi?id=43146
...but I think this is where we have to start.
We need to canonicalize/optimize sequences of shift and logic to ease
pattern matching for things like bswap and improve perf in general.
But without the artificial limit of '!LegalTypes' (early combining),
there are a lot of test diffs, and not all are good.
In the minimal tests added for this proposal, x86 should have better
throughput in all cases. AArch64 is neutral for scalar tests because
it can fold shifts into bitwise logic ops.
There are 3 shift opcodes and 3 logic opcodes for a total of 9 possible patterns:
https://rise4fun.com/Alive/VlIhttps://rise4fun.com/Alive/n1mhttps://rise4fun.com/Alive/1Vn
Differential Revision: https://reviews.llvm.org/D67021
llvm-svn: 370617
These were never enabled correctly and are causing other problems. Taking them
out for the moment, whilst we work on the issues.
This reverts r370329.
llvm-svn: 370607
The sequence between the function call and the asm start
may change without affecting what this test is looking for,
but we should have a better idea about what that sequence
looks like.
llvm-svn: 370518
Alina Sbirlea