This changes the lowering of saddsat and ssubsat so that instead of
using:
r,o = saddo x, y
c = setcc r < 0
s = c ? INTMAX : INTMIN
ret o ? s : r
into using asr and xor to materialize the INTMAX/INTMIN constants:
r,o = saddo x, y
s = ashr r, BW-1
x = xor s, INTMIN
ret o ? x : r
https://alive2.llvm.org/ce/z/TYufgD
This seems to reduce the instruction count in most testcases across most
architectures. X86 has some custom lowering added to compensate for
cases where it can increase instruction count.
Differential Revision: https://reviews.llvm.org/D105853
This enables subreg liveness in the arm backend when MVE is present,
which allows the register allocator to detect when subregister are
alive/dead, compared to only acting on full registers. This can helps
produce better code on MVE with the way MQPR registers are made up of
SPR registers, but is especially helpful for MQQPR and MQQQQPR
registers, where there are very few "registers" available and being able
to split them up into subregs can help produce much better code.
Differential Revision: https://reviews.llvm.org/D107642
Similar to the MQPR register class as the MVE equivalent to QPR, this
adds MQQPR and MQQQQPR register classes for the MVE equivalents of QQPR
and QQQQPR registers. The MVE MQPR seemed have worked out quite well,
and adding MQQPR and MQQQQPR allows us to a little more accurately
specify the number of registers, calculating register pressure limits a
little better.
Differential Revision: https://reviews.llvm.org/D107463
Currently isReallyTriviallyReMaterializableGeneric() implementation
prevents rematerialization on any virtual register use on the grounds
that is not a trivial rematerialization and that we do not want to
extend liveranges.
It appears that LRE logic does not attempt to extend a liverange of
a source register for rematerialization so that is not an issue.
That is checked in the LiveRangeEdit::allUsesAvailableAt().
The only non-trivial aspect of it is accounting for tied-defs which
normally represent a read-modify-write operation and not rematerializable.
The test for a tied-def situation already exists in the
/CodeGen/AMDGPU/remat-vop.mir,
test_no_remat_v_cvt_f32_i32_sdwa_dst_unused_preserve.
The change has affected ARM/Thumb, Mips, RISCV, and x86. For the targets
where I more or less understand the asm it seems to reduce spilling
(as expected) or be neutral. However, it needs a review by all targets'
specialists.
Differential Revision: https://reviews.llvm.org/D106408
Given a constant operand, the MVE and DAGCombine combines could fight,
each redistributing in the opposite order. Add a guard to the MVE
vecreduce distribution to prevent that.
That seems this test does not check what was stated in the
comment anymore. Just switch to generated checks.
Differential Revision: https://reviews.llvm.org/D107590
This attempts to make more of RDA aware of potentially overlapping
subregisters. Some of this was already in place, with it iterating
through MCRegUnitIterators. This also replaces calls to
LiveRegs.contains(..) with !LiveRegs.available(..), and updates the
isValidRegUseOf and isValidRegDefOf to search subregs.
Differential Revision: https://reviews.llvm.org/D107351
If the block target for a WLSTP instruction is known to be out of range,
and cannot be fixed by the ARMBlockPlacementPass, we can relax it to a
DLSTP (and cmp/branch) to still allow the creation of tail predicated
loops. That is what this patch does, adding extra revert code to the
fallback path of ARMBlockPlacementPass.
Due to the code produced when reverting, this creates a DLSTP between a
Bcc and a Br. As a DLS isn't necessarily a terminator we need to split
the block to move the DLS/Br into.
Differential Revision: https://reviews.llvm.org/D104709
Currently, the default alignment is much larger than the actual size of
the vector in memory. Fix this to use a sane default.
For SVE, temporarily remove lowering of load/store operations for
predicates with less than 16 elements. The layout the backend was
assuming for SVE predicates with less than 16 elements doesn't agree
with the frontend. More work probably needs to be done here.
This change is, strictly speaking, not backwards-compatible at the
bitcode level. But probably nobody is actually depending on that; i1
vectors in memory are rare, and the code that does use them probably
ends up forcing the alignment to something sane anyway. If we think
this is a concern, I can restrict this to scalable vectors for now
(where it's actually causing issues for me at the moment).
Differential Revision: https://reviews.llvm.org/D88994
This distributes reductions based on the relative offset of loads, if
one is found from their operands. Given chains of reductions this will
then sort them in ascending load order, which in turn can help simple
prefetches latch on to increasing strides more easily.
Differential Revision: https://reviews.llvm.org/D106569
This adds a combine for adds of reductions, distributing them so that
they occur sequentially to enable better use of accumulating VADDVA
instructions. It combines:
add(X, add(vecreduce(Y), vecreduce(Z))) ->
add(add(X, vecreduce(Y)), vecreduce(Z))
and
add(add(A, reduce(B)), add(C, reduce(D))) ->
add(add(add(A, C), reduce(B)), reduce(D))
These together distribute the add's so that more reductions can be
selected to VADDVA.
Differential Revision: https://reviews.llvm.org/D106532
Under MVE we can use VADDV/VADDVA's to perform integer add reductions,
so it can be beneficial to use more reductions than summing subvectors
and reducing once. Especially for VMLAV/VMLAVA the mul can be
incorporated into the reduction, producing less instructions.
Some of the test cases currently get larger due to extra integer adds,
but will be improved in a followup patch.
Differential Revision: https://reviews.llvm.org/D106531
These are generated from SLP vectorization, for example
https://godbolt.org/z/ebxdPh1Kz. These backend tests
show cases that we can produce better code for.
This implements the isLoadFromStackSlot and isStoreToStackSlot for MVE
MVE_VSTRWU32 and MVE_VLDRWU32 functions. They behave the same as many
other loads/stores, expecting a FI in Op1 and zero offset in Op2. At the
same time this alters VLDR_P0_off and VSTR_P0_off to use the same code
too, as they too should be returning VPR in Op0, take a FI in Op1 and
zero offset in Op2.
Differential Revision: https://reviews.llvm.org/D106797
The register class required for some MVE loads/stores is more
constrained than the register we use when creating postinc. Make sure we
constrain the register class to keep the code correct.
This relaxes the VMLAV and VADDV reduction recognition code to handle
smaller than legal types, extending them as needed. That was already
handled for some reductions, this extends it to more types in a more
generic way. If a smaller than legal value is found it is extended to
the legal type as needed.
Differential Revision: https://reviews.llvm.org/D106051
Corollary to 1113e06821 this allows us to
match gather that dont produce a full vector width results. They use an
extended gather which is truncated back to the original type.
This patch makes vector spills valid for tail predication when all loads
from the same stack slot are within the loop
Differential Revision: https://reviews.llvm.org/D105443
We have a DAG combine for recognizing the sequence of nodes that make up
an MVE VQDMULH, but only currently handles specifically legal types.
This patch expands that to other power-2 vector types. For smaller than
legal types this means any_extending the type and casting it to a legal
type, using a VQDMULH where we only use some of the lanes. The result is
sign extended back to the original type, to properly set the invalid
lanes. Larger than legal types are split into chunks with extracts and
concat back together.
Differential Revision: https://reviews.llvm.org/D105814
For i64 reductions we currently try and convert add(VMLALV(X, Y), B) to
VMLALVA(B, X, Y), incorporating the addition into the VMLALVA. If we
have an add of an existing VMLALVA, this patch pushes the add up above
the VMLALVA so that it may potentially be simplified further, for
example being folded into another VMLALV.
Differential Revision: https://reviews.llvm.org/D105686
MVE does not have a VMLALV instruction that can perform v16i8 -> i64
reductions, like it does for v8i16->i64 and v4i32->i64 reductions. That
means that the pattern to create them will be spilt up by type
legalization, creating a lot of instructions.
This extends the patterns for matching i64 reductions a little to handle
the v16i8->i64 case. We need to turn them into a pair of v8i16->i64
VMLALVs that each perform half of the reduction and are summed together
(so the later is a VMLALVA). The order of the lanes does not matter for
the reduction so we generate a MVEEXT for the extension, that will
either be folded into a extending load or can be optimized to a
VREV/VMOVL. Some of the resulting codegen isn't optimal, but will be
improved in a later patch.
Differential Revision: https://reviews.llvm.org/D105680
Similar to D91921 (and D104515) this introduces two MVESEXT and MVEZEXT
nodes that larger-than-legal sext and zext are lowered to. These either
get optimized away or end up becoming a series of stack loads/store, in
order to perform the extending whilst keeping the order of the lanes
correct. They are generated from v8i16->v8i32, v16i8->v16i16 and
v16i8->v16i32 extends, potentially with a intermediate extend for the
larger v16i8->v16i32 extend. A number of combines have been added for
obvious cases that come up in tests, notably MVEEXT of shuffles. More
may be needed in the future, but this seems to cover most of the cases
that come up in the tests.
Differential Revision: https://reviews.llvm.org/D105090
Much like fixed-point to floating-point conversion, the converse can
also be transformed into a fixed-point VCVT. This patch transforms
multiplications of floating point numbers by 2^n into a VCVT_fix. The
exception is that a float to fixed conversion with 1 fractional bit
ends up being an FADD (FADD(x, x) emulates FMUL(x, 2)) rather than an FMUL so there is a special case for that. This patch also moves the code from https://reviews.llvm.org/D103903 into a separate function as fixed to float and float to fixed are very similar.
Differential Revision: https://reviews.llvm.org/D104793
This will currently accept the old number of bytes syntax, and convert
it to a scalar. This should be removed in the near future (I think I
converted all of the tests already, but likely missed a few).
Not sure what the exact syntax and policy should be. We can continue
printing the number of bytes for non-generic instructions to avoid
test churn and only allow non-scalar types for generic instructions.
This will currently print the LLT in parentheses, but accept parsing
the existing integers and implicitly converting to scalar. The
parentheses are a bit ugly, but the parser logic seems unable to deal
without either parentheses or some keyword to indicate the start of a
type.
This adds a small fold for extract (ARM_BUILD_VECTOR) to fold to the
original node. This can help simplify the resulting codegen in some
cases.
Differential Revision: https://reviews.llvm.org/D104860
This adds another small fold for extract of a vdup, between a i32 and a
f32, converting to a BITCAST. This allows some extra folding to happen,
simplifying the resulting code.
Differential Revision: https://reviews.llvm.org/D104857
The MVETRUNC node truncates two wide vectors to a single vector with
narrower elements. This is usually lowered to a series of extract/insert
elements, going via GPR registers. This patch changes that to instead
use a pair of truncating stores and a stack reload. This cuts down the
number of instructions at the expense of some stack space.
Differential Revision: https://reviews.llvm.org/D104515
Currently, when encountering store(trunc(..)) where the trunc is double
a legal vector lenth in MVE, we spilt the node into two different stores
each performing half of the trunc from the wider type. This works well
for efficiently lowering wider than legal types, else the trunc becomes
a series of individual lane moves. Unfortunately this splitting is
currently one of the first combines attempted, so can happen before any
other combines which might be more preferable.
This patch instead introduces the concept of a MVETRUNC ISel node that
the trunk is initially lowered to, to keep it intact as a single item as
opposed to splitting it up. This allows us to push the store(trunc(..))
combine later, allowing other optimisations to potentially happen on the
trunc first. The store(trunc(..)) splitting can then be done later in
the legalisation period if needed, or else fall back to a buildvector as
before.
This can also be used in the future to lower to loads/stores, as opposed
to the more expensive lane extracts/inserts. Some extra combines are
added to keep all the existing tests happy.
Differential Revision: https://reviews.llvm.org/D91921
As a minor adjustment to the existing lowering of offset scatters, this
extends any smaller-than-legal vectors into full vectors using a zext,
so that the truncating scatters can be used. Due to the way MVE
legalizes the vectors this should be cheap in most situations, and will
prevent the vector from being scalarized.
Differential Revision: https://reviews.llvm.org/D103704
Based ontop of D104598, which is a NFCI-ish refactoring.
Here, a restriction, that only empty blocks can be merged, is lifted.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D104597
Conversion from a fixed-point number to a floating-point number is done by
multiplying the fixed-point number by 2^(-n) where n is the number of
fractional bits. Currently this is lowered to a vcvt
(integer to floating-point) then a vmul, but it can instead be lowered
directly to a vcvt (fixed-point to floating-point). This patch enables
such transformations as long as the multiplication factor is a power of 2.
Differential Revision: https://reviews.llvm.org/D103903
These all (and some others) are being affected by D104597,
but they are manually-written, which rather complicates
checking the effect that change has on them.
This can be seen as a follow up to commit 0ee439b705,
that changed the second argument of __powidf2, __powisf2 and
__powitf2 in compiler-rt from si_int to int. That was to align with
how those runtimes are defined in libgcc.
One thing that seem to have been missing in that patch was to make
sure that the rest of LLVM also handle that the argument now depends
on the size of int (not using the si_int machine mode for 32-bit).
When using __builtin_powi for a target with 16-bit int clang crashed.
And when emitting libcalls to those rtlib functions, typically when
lowering @llvm.powi), the backend would always prepare the exponent
argument as an i32 which caused miscompiles when the rtlib was
compiled with 16-bit int.
The solution used here is to use an overloaded type for the second
argument in @llvm.powi. This way clang can use the "correct" type
when lowering __builtin_powi, and then later when emitting the libcall
it is assumed that the type used in @llvm.powi matches the rtlib
function.
One thing that needed some extra attention was that when vectorizing
calls several passes did not support that several arguments could
be overloaded in the intrinsics. This patch allows overload of a
scalar operand by adding hasVectorInstrinsicOverloadedScalarOpd, with
an entry for powi.
Differential Revision: https://reviews.llvm.org/D99439
Under MVE v4f32 and v8f16 vectors should be using v4i1/v8i1 predicates
for the setcc result type, as they have predicated registers for those
types. Setting this correctly prevents some inefficient optimizations
from happening.
David Green