Summary:
This patch adds intrinsics for the following MVE instructions:
* VABAV
* VMLADAV, VMLSDAV
* VMLALDAV, VMLSLDAV
* VRMLALDAVH, VRMLSLDAVH
Each of the above 4 groups has a corresponding new LLVM IR intrinsic,
since the instructions cannot be easily represented using
general-purpose IR operations.
Reviewers: simon_tatham, ostannard, dmgreen, MarkMurrayARM
Reviewed By: MarkMurrayARM
Subscribers: merge_guards_bot, kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71062
Summary:
This fills in the remaining shift operations that take a single vector
input and an immediate shift count: the `vqshl`, `vqshlu`, `vrshr` and
`vshll[bt]` families.
`vshll[bt]` (which shifts each input lane left into a double-width
output lane) is the most interesting one. There are separate MC
instruction ids for shifting by exactly the input lane width and
shifting by less than that, because the instruction encoding is so
completely different for the lane-width special case. So I had to
write two sets of patterns to match based on the immediate shift
count, which involved adding a ComplexPattern matcher to avoid the
general-case pattern accidentally matching the special case too. For
that family I've made sure to add an llc codegen test for both
versions of each instruction.
I'm experimenting with a new strategy for parametrising the isel
patterns for all these instructions: adding extra fields to the
relevant `Instruction` subclass itself, which are ignored by the
Tablegen backends that generate the MC data, but can be retrieved from
each instance of that instruction subclass when it's passed as a
template parameter to the multiclass that generates its isel patterns.
A nice effect of that is that I can fill in those informational fields
using `let` blocks, rather than having to type them out once per
instruction at `defm` time.
(As a result, quite a lot of existing instruction `def`s are
reindented by this patch, so it's clearer to read with whitespace
changes ignored.)
Reviewers: dmgreen, MarkMurrayARM, miyuki, ostannard
Reviewed By: MarkMurrayARM
Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71458
This has two main effects:
- Optimizes debug info size by saving 221.86 MB of obj file size in a
Windows optimized+debug build of 'all'. This is 3.03% of 7,332.7MB of
object file size.
- Incremental step towards decoupling target intrinsics.
The enums are still compact, so adding and removing a single
target-specific intrinsic will trigger a rebuild of all of LLVM.
Assigning distinct target id spaces is potential future work.
Part of PR34259
Reviewers: efriedma, echristo, MaskRay
Reviewed By: echristo, MaskRay
Differential Revision: https://reviews.llvm.org/D71320
Summary:
This patch refactors instruction selection of the complex vector
addition, multiplication and multiply-add intrinsics, so that it is
now based on TableGen patterns rather than C++ code.
It also changes the first parameter (halving vs non-halving) of the
arm_mve_vcaddq IR intrinsic to match the corresponding instruction
encoding, hence it requires some changes in the tests.
The patch addresses David's comment in https://reviews.llvm.org/D71190
Reviewers: dmgreen, ostannard, simon_tatham, MarkMurrayARM
Reviewed By: dmgreen
Subscribers: merge_guards_bot, kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71245
Summary:
This patch adds intrinsics for the following MVE instructions:
* VCADD, VHCADD
* VCMUL
* VCMLA
Each of the above 3 groups has a corresponding new LLVM IR intrinsic.
Reviewers: simon_tatham, MarkMurrayARM, ostannard, dmgreen
Reviewed By: MarkMurrayARM
Subscribers: merge_guards_bot, kristof.beyls, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D71190
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
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
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
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
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
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
Summary:
This clang-tidy check is looking for unsigned integer variables whose initializer
starts with an implicit cast from llvm::Register and changes the type of the
variable to llvm::Register (dropping the llvm:: where possible).
Partial reverts in:
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
X86FixupLEAs.cpp - Some functions return unsigned and arguably should be MCRegister
X86FrameLowering.cpp - Some functions return unsigned and arguably should be MCRegister
HexagonBitSimplify.cpp - Function takes BitTracker::RegisterRef which appears to be unsigned&
MachineVerifier.cpp - Ambiguous operator==() given MCRegister and const Register
PPCFastISel.cpp - No Register::operator-=()
PeepholeOptimizer.cpp - TargetInstrInfo::optimizeLoadInstr() takes an unsigned&
MachineTraceMetrics.cpp - MachineTraceMetrics lacks a suitable constructor
Manual fixups in:
ARMFastISel.cpp - ARMEmitLoad() now takes a Register& instead of unsigned&
HexagonSplitDouble.cpp - Ternary operator was ambiguous between unsigned/Register
HexagonConstExtenders.cpp - Has a local class named Register, used llvm::Register instead of Register.
PPCFastISel.cpp - PPCEmitLoad() now takes a Register& instead of unsigned&
Depends on D65919
Reviewers: arsenm, bogner, craig.topper, RKSimon
Reviewed By: arsenm
Subscribers: RKSimon, craig.topper, lenary, aemerson, wuzish, jholewinski, MatzeB, qcolombet, dschuff, jyknight, dylanmckay, sdardis, nemanjai, jvesely, wdng, nhaehnle, sbc100, jgravelle-google, kristof.beyls, hiraditya, aheejin, kbarton, fedor.sergeev, javed.absar, asb, rbar, johnrusso, simoncook, apazos, sabuasal, niosHD, jrtc27, MaskRay, zzheng, edward-jones, atanasyan, rogfer01, MartinMosbeck, brucehoult, the_o, tpr, PkmX, jocewei, jsji, Petar.Avramovic, asbirlea, Jim, s.egerton, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65962
llvm-svn: 369041
This adds pre- and post- increment and decrements for MVE loads and stores. It
uses the builtin pre and post load/store detection, unlike Neon. Loads are
selected with the code in tryT2IndexedLoad, stores are selected with tablegen
patterns. The immediates have a +/-7bit range, multiplied by the size of the
element.
Differential Revision: https://reviews.llvm.org/D63840
llvm-svn: 368305
While lowering test.set.loop.iterations, it wasn't checked how the
brcond was using the result and so the wls could branch to the loop
preheader instead of not entering it. The same was true for
loop.decrement.reg.
So brcond and br_cc and now lowered manually when using the hwloop
intrinsics. During this we now check whether the result has been
negated and whether we're using SETEQ or SETNE and 0 or 1. We can
then figure out which basic block the WLS and LE should be targeting.
Differential Revision: https://reviews.llvm.org/D64616
llvm-svn: 366809
Backend changes to enable WLS/LE low-overhead loops for armv8.1-m:
1) Use TTI to communicate to the HardwareLoop pass that we should try
to generate intrinsics that guard the loop entry, as well as setting
the loop trip count.
2) Lower the BRCOND that uses said intrinsic to an Arm specific node:
ARMWLS.
3) ISelDAGToDAG the node to a new pseudo instruction:
t2WhileLoopStart.
4) Add support in ArmLowOverheadLoops to handle the new pseudo
instruction.
Differential Revision: https://reviews.llvm.org/D63816
llvm-svn: 364733
The current implementation of ThumbRegisterInfo::saveScavengerRegister
is bad for two reasons: one, it's buggy, and two, it blocks using R12
for other optimizations. So this patch gets rid of it, and adds the
necessary support for using an ordinary emergency spill slot on Thumb1.
(Specifically, I think saveScavengerRegister was broken by r305625, and
nobody noticed for two years because the codepath is almost never used.
The new code will also probably not be used much, but it now has better
tests, and if we fail to emit a necessary emergency spill slot we get a
reasonable error message instead of a miscompile.)
A rough outline of the changes in the patch:
1. Gets rid of ThumbRegisterInfo::saveScavengerRegister.
2. Modifies ARMFrameLowering::determineCalleeSaves to allocate an
emergency spill slot for Thumb1.
3. Implements useFPForScavengingIndex, so the emergency spill slot isn't
placed at a negative offset from FP on Thumb1.
4. Modifies the heuristics for allocating an emergency spill slot to
support Thumb1. This includes fixing ExtraCSSpill so we don't try to
use "lr" as a substitute for allocating an emergency spill slot.
5. Allocates a base pointer in more cases, so the emergency spill slot
is always accessible.
6. Modifies ARMFrameLowering::ResolveFrameIndexReference to compute the
right offset in the new cases where we're forcing a base pointer.
7. Ensures we never generate a load or store with an offset outside of
its frame object. This makes the heuristics more straightforward.
8. Changes Thumb1 prologue and epilogue emission so it never uses
register scavenging.
Some of the changes to the emergency spill slot heuristics in
determineCalleeSaves affect ARM/Thumb2; hopefully, they should allow
the compiler to avoid allocating an emergency spill slot in cases
where it isn't necessary. The rest of the changes should only affect
Thumb1.
Differential Revision: https://reviews.llvm.org/D63677
llvm-svn: 364490
This provides the low-level support to start using MVE vector types in
LLVM IR, loading and storing them, passing them to __asm__ statements
containing hand-written MVE vector instructions, and *if* you have the
hard-float ABI turned on, using them as function parameters.
(In the soft-float ABI, vector types are passed in integer registers,
and combining all those 32-bit integers into a q-reg requires support
for selection DAG nodes like insert_vector_elt and build_vector which
aren't implemented yet for MVE. In fact I've also had to add
`arm_aapcs_vfpcc` to a couple of existing tests to avoid that
problem.)
Specifically, this commit adds support for:
* spills, reloads and register moves for MVE vector registers
* ditto for the VPT predication mask that lives in VPR.P0
* make all the MVE vector types legal in ISel, and provide selection
DAG patterns for BITCAST, LOAD and STORE
* make loads and stores of scalar FP types conditional on
`hasFPRegs()` rather than `hasVFP2Base()`. As a result a few
existing tests needed their llc command lines updating to use
`-mattr=-fpregs` as their method of turning off all hardware FP
support.
Reviewers: dmgreen, samparker, SjoerdMeijer
Subscribers: javed.absar, kristof.beyls, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60708
llvm-svn: 364329
Introduce three pseudo instructions to be used during DAG ISel to
represent v8.1-m low-overhead loops. One maps to set_loop_iterations
while loop_decrement_reg is lowered to two, so that we can separate
the decrement and branching operations. The pseudo instructions are
expanded pre-emission, where we can still decide whether we actually
want to generate a low-overhead loop, in a new pass:
ARMLowOverheadLoops. The pass currently bails, reverting to an sub,
icmp and br, in the cases where a call or stack spill/restore happens
between the decrement and branching instructions, or if the loop is
too large.
Differential Revision: https://reviews.llvm.org/D63476
llvm-svn: 364288
Those two subtarget features were awkward because their semantics are
reversed: each one indicates the _lack_ of support for something in
the architecture, rather than the presence. As a consequence, you
don't get the behavior you want if you combine two sets of feature
bits.
Each SubtargetFeature for an FP architecture version now comes in four
versions, one for each combination of those options. So you can still
say (for example) '+vfp2' in a feature string and it will mean what
it's always meant, but there's a new string '+vfp2d16sp' meaning the
version without those extra options.
A lot of this change is just mechanically replacing positive checks
for the old features with negative checks for the new ones. But one
more interesting change is that I've rearranged getFPUFeatures() so
that the main FPU feature is appended to the output list *before*
rather than after the features derived from the Restriction field, so
that -fp64 and -d32 can override defaults added by the main feature.
Reviewers: dmgreen, samparker, SjoerdMeijer
Subscribers: srhines, javed.absar, eraman, kristof.beyls, hiraditya, zzheng, Petar.Avramovic, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D60691
llvm-svn: 361845
This takes sequences like "mov r4, sp; str r0, [r4]", and optimizes them
to something like "str r0, [sp]".
For regular stack variables, this optimization was already implemented:
we lower loads and stores using frame indexes, which are expanded later.
However, when constructing a call frame for a call with more than four
arguments, the existing optimization doesn't apply. We need to use
stores which are actually relative to the current value of sp, and don't
have an associated frame index.
This patch adds a special case to handle that construct. At the DAG
level, this is an ISD::STORE where the address is a CopyFromReg from SP
(plus a small constant offset).
This applies only to Thumb1: in Thumb2 or ARM mode, a regular store
instruction can access SP directly, so the COPY gets eliminated by
existing code.
The change to ARMDAGToDAGISel::SelectThumbAddrModeSP is a related
cleanup: we shouldn't pretend that it can select anything other than
frame indexes.
Differential Revision: https://reviews.llvm.org/D59568
llvm-svn: 356601
The isScaledConstantInRange function takes upper and lower bounds which are
checked after dividing by the scale, so the bounds checks for half, single and
double precision should all be the same. Previously, we had wrong bounds checks
for half precision, so selected an immediate the instructions can't actually
represent.
Differential revision: https://reviews.llvm.org/D58822
llvm-svn: 355305
This patch accompanies the RFC posted here:
http://lists.llvm.org/pipermail/llvm-dev/2018-October/127239.html
This patch adds a new CallBr IR instruction to support asm-goto
inline assembly like gcc as used by the linux kernel. This
instruction is both a call instruction and a terminator
instruction with multiple successors. Only inline assembly
usage is supported today.
This also adds a new INLINEASM_BR opcode to SelectionDAG and
MachineIR to represent an INLINEASM block that is also
considered a terminator instruction.
There will likely be more bug fixes and optimizations to follow
this, but we felt it had reached a point where we would like to
switch to an incremental development model.
Patch by Craig Topper, Alexander Ivchenko, Mikhail Dvoretckii
Differential Revision: https://reviews.llvm.org/D53765
llvm-svn: 353563
In many places in the backend, we like to know whether we're
optimising for code size and this is performed by checking the
current machine function attributes. A subtarget is created on a
per-function basis, so it's possible to know when we're compiling for
code size on construction so record this in the new object.
Differential Revision: https://reviews.llvm.org/D57812
llvm-svn: 353501
Constants can also be materialised using the negated value and a MVN, and this
case seem to have been missed for Thumb2. To check the constant materialisation
costs, we now call getT2SOImmVal twice, once for the original constant and then
also for its negated value, and this function checks if the constant can both
be splatted or rotated.
This was revealed by a test that optimises for minsize: instead of a LDR
literal pool load and having a literal pool entry, just a MVN with an immediate
is smaller (and also faster).
Differential Revision: https://reviews.llvm.org/D57327
llvm-svn: 352737
This attempts to optimise negative values used in load/store operands
a little. We currently try to selct them as rr, materialising the
negative constant using a MOV/MVN pair. This instead selects ri with
an immediate of 0, forcing the add node to become a simpler sub.
Differential Revision: https://reviews.llvm.org/D57121
llvm-svn: 352475
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
`MachineMemOperand` pointers attached to `MachineSDNodes` and instead
have the `SelectionDAG` fully manage the memory for this array.
Prior to this change, the memory management was deeply confusing here --
The way the MI was built relied on the `SelectionDAG` allocating memory
for these arrays of pointers using the `MachineFunction`'s allocator so
that the raw pointer to the array could be blindly copied into an
eventual `MachineInstr`. This creates a hard coupling between how
`MachineInstr`s allocate their array of `MachineMemOperand` pointers and
how the `MachineSDNode` does.
This change is motivated in large part by a change I am making to how
`MachineFunction` allocates these pointers, but it seems like a layering
improvement as well.
This would run the risk of increasing allocations overall, but I've
implemented an optimization that should avoid that by storing a single
`MachineMemOperand` pointer directly instead of allocating anything.
This is expected to be a net win because the vast majority of uses of
these only need a single pointer.
As a side-effect, this makes the API for updating a `MachineSDNode` and
a `MachineInstr` reasonably different which seems nice to avoid
unexpected coupling of these two layers. We can map between them, but we
shouldn't be *surprised* at where that occurs. =]
Differential Revision: https://reviews.llvm.org/D50680
llvm-svn: 339740
LLVM normally prefers to minimize the number of bits set in an AND
immediate, but that doesn't always match the available ARM instructions.
In Thumb1 mode, prefer uxtb or uxth where possible; otherwise, prefer
a two-instruction sequence movs+ands or movs+bics.
Some potential improvements outlined in
ARMTargetLowering::targetShrinkDemandedConstant, but seems to work
pretty well already.
The ARMISelDAGToDAG fix ensures we don't generate an invalid UBFX
instruction due to a larger-than-expected mask. (It's orthogonal, in
some sense, but as far as I can tell it's either impossible or nearly
impossible to reproduce the bug without this change.)
According to my testing, this seems to consistently improve codesize by
a small amount by forming bic more often for ISD::AND with an immediate.
Differential Revision: https://reviews.llvm.org/D50030
llvm-svn: 339472
We don't ever check these again (unless you're using
-fno-integrated-as), so make sure the extracted bits are well-defined.
I don't think it's possible to trigger any of the assertions on trunk,
but it's difficult to prove. (The first one depends on DAGCombine to
minimize the number of set bits in AND masks; I think the others are
mathematically impossible to hit.)
llvm-svn: 335931
This patch adds support for the q versions of the dup
(load-to-all-lanes) NEON intrinsics, such as vld2q_dup_f16() for
example.
Currently, non-q versions of the dup intrinsics are implemented
in clang by generating IR that first loads the elements of the
structure into the first lane with the lane (to-single-lane)
intrinsics, and then propagating it other lanes. There are at
least two problems with this approach. First, there are no
double-spaced to-single-lane byte-element instructions. For
example, there is no such instruction as 'vld2.8 { d0[0], d2[0]
}, [r0]'. That means we cannot rely on the to-single-lane
intrinsics and instructions to implement the q versions of the
dup intrinsics. Note that to-all-lanes instructions do support
all sizes of data items, including bytes.
The second problem with the current approach is that we need a
separate vdup instruction to propagate the structure to each
lane. So for vld4q_dup_f16() we would need four vdup instructions
in addition to the initial vld instruction.
This patch introduces dup LLVM intrinsics and reworks handling of
the currently supported (non-q) NEON dup intrinsics to expand
them into those LLVM intrinsics, thus eliminating the need for
using to-single-lane intrinsics and instructions.
Additionally, this patch adds support for u64 and s64 dup NEON
intrinsics. These are marked as Arch64-only in the ARM NEON
Reference, but it seems there are no reasons to not support them
in AArch32 mode. Please correct, if that is wrong.
That's what we generate with this patch applied:
vld2q_dup_f16:
vld2.16 {d0[], d2[]}, [r0]
vld2.16 {d1[], d3[]}, [r0]
vld3q_dup_f16:
vld3.16 {d0[], d2[], d4[]}, [r0]
vld3.16 {d1[], d3[], d5[]}, [r0]
vld4q_dup_f16:
vld4.16 {d0[], d2[], d4[], d6[]}, [r0]
vld4.16 {d1[], d3[], d5[], d7[]}, [r0]
Differential Revision: https://reviews.llvm.org/D48439
llvm-svn: 335733
Thumb has more 16-bit encoding space dedicated to ADD than ORR, allowing both a
3-address encoding and a wider range of immediates. So, particularly when
optimizing for code size (but it doesn't make things worse elsewhere) it's
beneficial to select an OR operation to an ADD if we know overflow won't occur.
This is made even better by LLVM's penchant for putting operations in canonical
form by converting the other way.
llvm-svn: 335119
We currently support them only in AArch64. The NEON Reference,
however, says they are 'ARMv7, ARMv8' intrinsics.
Differential Revision: https://reviews.llvm.org/D47447
llvm-svn: 334361
It looks like this got left in by accident in r289794; I can't think of
any reason this check would be necessary. (Maybe it was meant to be a
check that the AND has one use? But we check that a few lines earlier.)
Differential Revision: https://reviews.llvm.org/D47921
llvm-svn: 334322
We currently support them only in AArch64. The NEON Reference,
however, says they are 'ARMv7, ARMv8' intrinsics.
Differential Revision: https://reviews.llvm.org/D47120
llvm-svn: 333825
We currently support them only in AArch64. The NEON Reference,
however, says they are 'ARMv7, ARMv8' intrinsics.
Differential Revision: https://reviews.llvm.org/D47121
llvm-svn: 333819
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46290
llvm-svn: 331272
This is the groundwork for adding the Armv8.2-A FP16 vector intrinsics, which
uses v4f16 and v8f16 vector operands and return values. All the moving parts
are tested with two intrinsics, a 1-operand v8f16 and a 2-operand v4f16
intrinsic. In a follow-up patch the rest of the intrinsics and tests will be
added.
Differential Revision: https://reviews.llvm.org/D44538
llvm-svn: 327839
r327171 "Improve Dependency analysis when doing multi-node Instruction Selection"
r328170 "[DAG] Enforce stricter NodeId invariant during Instruction selection"
Reverting patch as NodeId invariant change is causing pathological
increases in compile time on PPC
llvm-svn: 327197
Instruction Selection makes use of the topological ordering of nodes
by node id (a node's operands have smaller node id than it) when doing
cycle detection. During selection we may violate this property as a
selection of multiple nodes may induce a use dependence (and thus a
node id restriction) between two unrelated nodes. If a selected node
has an unselected successor this may allow us to miss a cycle in
detection an invalid selection.
This patch fixes this by marking all unselected successors of a
selected node have negated node id. We avoid pruning on such negative
ids but still can reconstruct the original id for pruning.
In-tree targets have been updated to replace DAG-level replacements
with ISel-level ones which enforce this property.
This preemptively fixes PR36312 before triggering commit r324359 relands
Reviewers: craig.topper, bogner, jyknight
Subscribers: arsenm, nhaehnle, javed.absar, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D43198
llvm-svn: 327170
Code generation of VLD3, VLD4, VST3 and VST4 with register writeback is
broken due to 2 separate bugs:
1) VLD1d64TPseudoWB_register and VLD1d64QPseudoWB_register are missing
rules to expand them to non pseudo MIR. These are selected for
ARMISD::VLD3_UPD/VLD4_UPD with v1i64 vectors in SelectVLD.
2) Selection of the right VLD/VST instruction is broken for load and
store of 3 and 4 v1i64 vectors. SelectVLD and SelectVST are called
with MIR opcode for fixed writeback (ie increment is access size)
and call getVLDSTRegisterUpdateOpcode() to select an opcode with
register writeback if base register update is of a different size.
Since getVLDSTRegisterUpdateOpcode() only knows about
VLD1/VLD2/VST1/VST2 the call is currently conditional on the number
of element in the vector.
However, VLD1/VST1 is selected by SelectVLD/SelectVST's caller for
load and stores of 3 or 4 v1i64 vectors. Therefore the opcode is not
updated which later lead to a fixed writeback instruction being
constructed with an extra operand for the register writeback.
This patch addresses the two issues as follows:
- it adds the necessary mapping from VLD1d64TPseudoWB_register and
VLD1d64QPseudoWB_register to VLD1d64Twb_register and
VLD1d64Qwb_register respectively. Like for the existing _fixed
variants, the cost of these is bumped for unaligned access.
- it changes the logic in SelectVLD and SelectVSD to call isVLDfixed
and isVSTfixed respectively to decide whether the opcode should be
updated. It also reworks the logic and comments for pushing the
writeback offset operand and r0 operand to clarify the logic:
writeback offset needs to be pushed if it's a register writeback,
r0 needs to be pushed if not and the instruction is a
VLD1/VLD2/VST1/VST2.
Reviewers: rengolin, t.p.northover, samparker
Reviewed By: samparker
Patch by Thomas Preud'homme <thomas.preudhomme@arm.com>
Differential Revision: https://reviews.llvm.org/D42970
llvm-svn: 326570
This is the groundwork for Armv8.2-A FP16 code generation .
Clang passes and returns _Float16 values as floats, together with the required
bitconverts and truncs etc. to implement correct AAPCS behaviour, see D42318.
We will implement half-precision argument passing/returning lowering in the ARM
backend soon, but for now this means that this:
_Float16 sub(_Float16 a, _Float16 b) {
return a + b;
}
gets lowered to this:
define float @sub(float %a.coerce, float %b.coerce) {
entry:
%0 = bitcast float %a.coerce to i32
%tmp.0.extract.trunc = trunc i32 %0 to i16
%1 = bitcast i16 %tmp.0.extract.trunc to half
<SNIP>
%add = fadd half %1, %3
<SNIP>
}
When FullFP16 is *not* supported, we don't make f16 a legal type, and we get
legalization for "free", i.e. nothing changes and everything works as before.
And also f16 argument passing/returning is handled.
When FullFP16 is supported, we do make f16 a legal type, and have 2 places that
we need to patch up: f16 argument passing and returning, which involves minor
tweaks to avoid unnecessary code generation for some bitcasts.
As a "demonstrator" that this works for the different FP16, FullFP16, softfp
modes, etc., I've added match rules to the VSUB instruction description showing
that we can codegen this instruction from IR, but more importantly, also to
some conversion instructions. These conversions were causing issue before in
the FP16 and FullFP16 cases.
I've also added match rules to the VLDRH and VSTRH desriptions, so that we can
actually compile the entire half-precision sub code example above. This showed
that these loads and stores had the wrong addressing mode specified: AddrMode5
instead of AddrMode5FP16, which turned out not be implemented at all, so that
has also been added.
This is the minimal patch that shows all the different moving parts. In patch
2/3 I will add some efficient lowering of bitcasts, and in 2/3 I will add the
remaining Armv8.2-A FP16 instruction descriptions.
Thanks to Sam Parker and Oliver Stannard for their help and reviews!
Differential Revision: https://reviews.llvm.org/D38315
llvm-svn: 323512
All these headers already depend on CodeGen headers so moving them into
CodeGen fixes the layering (since CodeGen depends on Target, not the
other way around).
llvm-svn: 318490
This is a refactoring/cleanup of Arm `addrmode2` operand class. The patch
removes it completely.
Differential Revision: https://reviews.llvm.org/D39832
llvm-svn: 318291
Moves encoding (SYSm) information of banked registers to ARMSystemRegister.td,
where it rightly belongs and forms a single point of reference in the code.
Reviewed by: @fhahn, @rovka, @olista01
Differential Revision: https://reviews.llvm.org/D36219
llvm-svn: 309910
This patch cleans up and fixes issues in the M-Class system register handling:
1. It defines the system registers and the encoding (SYSm values) in one place:
a new ARMSystemRegister.td using SearchableTable, thereby removing the
hand-coded values which existed in multiple places.
2. Some system registers e.g. BASEPRI_MAX_NS which do not exist were being allowed!
Ref: ARMv6/7/8M architecture reference manual.
Reviewed by: @t.p.northover, @olist01, @john.brawn
Differential Revision: https://reviews.llvm.org/D35209
llvm-svn: 308456
Constants are crucial for code size in the ARM Thumb-1 instruction
set. The 16 bit instruction size often does not offer enough space
for immediate arguments. This means that additional instructions are
frequently used to load constants into registers. Since constants are
hoisted, this can lead to significant register spillage if they are
used multiple times in a single function. This can be avoided by
rematerialization, i.e. recomputing a constant instead of reloading
it from the stack. This patch fixes the rematerialization of literal
pool loads in the ARM Thumb instruction set.
Patch by Philip Ginsbach
Differential Revision: https://reviews.llvm.org/D33936
llvm-svn: 308004
When we replaced the multiplicand the destination node might already exist.
When that happens the original gets CSEd and deleted. However, it's actually
used as the offset so nonsense is produced.
Should fix PR32726.
llvm-svn: 301983
Before, we assumed that any ConstantInt offset was precisely the access width,
so we could use the "[rN]!" form. ISelLowering only ever created that kind, but
further simplification during combining could lead to unexpected constants and
incorrect codegen.
Should fix PR32658.
llvm-svn: 300878
This will become asan errors once the patch lands that poisons the
memory after free. The x86 change is a hack, but I don't see how to
solve this properly at the moment.
llvm-svn: 300867
Create nodes for smulwb and smulwt and move their selection from
DAGToDAG to DAG combine. smlawb and smlawt can then be selected
using tablegen. Added some helper functions to detect shift patterns
as well as a wrapper around SimplifyDemandBits. Added a couple of
extra tests.
Differential Revision: https://reviews.llvm.org/D30708
llvm-svn: 297716
ARMISD::ADD[CE] nodes, instead of the generic ISD::ADD[CE].
Summary:
This allows for some simplification because the combines
are no longer limited to just one go at the node before
it gets legalized into an ARM target-specific one.
Reviewers: jmolloy, rogfer01
Subscribers: aemerson, llvm-commits, rengolin
Differential Revision: https://reviews.llvm.org/D30401
llvm-svn: 297453
In the encoding of system registers in the M-class MSR instruction the mask bits
should be 2 for registers that don't take a _<bits> qualifier (the instruction
is unpredictable otherwise), and should also be 2 if the register takes a
_<bits> qualifier but it's not present as no _<bits> is an alias for _nzcvq.
Differential Revision: https://reviews.llvm.org/D29828
llvm-svn: 294762
Currently, there are substantial problems forming vld1_dup even if the
VDUP survives legalization. The lack of an actual node
leads to terrible results: not only can we not form post-increment vld1_dup
instructions, but we form scalar pre-increment and post-increment
loads which force the loaded value into a GPR. This patch fixes that
by combining the vdup+load into an ARMISD node before DAGCombine
messes it up.
Also includes a crash fix for vld2_dup (see testcase @vld2dupi8_postinc_variable).
Recommiting with fix to avoid forming vld1dup if the type of the load
doesn't match the type of the vdup (see
https://llvm.org/bugs/show_bug.cgi?id=31404).
Differential Revision: https://reviews.llvm.org/D27694
llvm-svn: 289972
This is essentially a recommit of r285893, but with a correctness fix. The
problem of the original commit was that this:
bic r5, r7, #31
cbz r5, .LBB2_10
got rewritten into:
lsrs r5, r7, #5
beq .LBB2_10
The result in destination register r5 is not the same and this is incorrect
when r5 is not dead. So this fix includes checking the uses of the AND
destination register. And also, compared to the original commit, some regression
tests didn't need changing anymore because of this extra check.
For completeness, this was the original commit message:
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more
efficient instruction selection if the bitmask is one consecutive sequence of
set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and
set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and
set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit
into the sign bit with one LSLS and change the condition query from NE/EQ to
MI/PL (we could also implement this by shifting into the carry bit and
branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower
zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two
16-bit instructions but can elide the CMP and doesn't require materializing a
complex immediate, so is also a win.
Differential Revision: https://reviews.llvm.org/D27761
llvm-svn: 289794
Currently, there are substantial problems forming vld1_dup even if the
VDUP survives legalization. The lack of an actual node
leads to terrible results: not only can we not form post-increment vld1_dup
instructions, but we form scalar pre-increment and post-increment
loads which force the loaded value into a GPR. This patch fixes that
by combining the vdup+load into an ARMISD node before DAGCombine
messes it up.
Also includes a crash fix for vld2_dup (see testcase @vld2dupi8_postinc_variable).
Differential Revision: https://reviews.llvm.org/D27694
llvm-svn: 289703
This recommits r281323, which was backed out for two reasons. One, a selfhost failure, and two, it apparently caused Chromium failures. Actually, the latter was a red herring. The log has expired from the former, but I suspect that was a red herring too (actually caused by another problematic patch of mine). Therefore reapplying, and will watch the bots like a hawk.
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 285893
UMAAL is a DSP instruction and it is not available on thumbv7m
(Cortex-M3) and thumbv6m (Cortex-M0+1) targets. Also fix wrong
CHECK prefix in longMAC.ll test.
Patch by Vadzim Dambrouski.
Differential Revision: https://reviews.llvm.org/D25890
llvm-svn: 285278
This is not a valid encoding - these instructions cannot do PC-relative addressing.
The underlying problem here is of whitelist in ARMISelDAGToDAG that unwraps ARMISD::Wrappers during addressing-mode selection. This didn't realise TargetConstantPool was actually possible, so didn't handle it.
llvm-svn: 283323
ldm and stm instructions always require 4-byte alignment on the pointer, but we
weren't checking this before trying to reduce code-size by replacing a
post-indexed load/store with them. Unfortunately, we were also dropping this
incormation in DAG ISel too, but that's easy enough to fix.
llvm-svn: 281893
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 281323
For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more efficient instruction selection if the bitmask is one consecutive sequence of set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).
1) If the bitmask touches the LSB, then we can remove all the upper bits and set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit into the sign bit with one LSLS and change the condition query from NE/EQ to MI/PL (we could also implement this by shifting into the carry bit and branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower zero bits of the mask.
1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two 16-bit instructions but can elide the CMP and doesn't require materializing a complex immediate, so is also a win.
llvm-svn: 281215
The CMPZ #0 disappears during peepholing, leaving just a tADDi3, tADDi8 or t2ADDri. This avoids having to materialize the expensive negative constant in Thumb-1, and allows a shrinking from a 32-bit CMN to a 16-bit ADDS in Thumb-2.
llvm-svn: 281040
This is a mechanical change of comments in switches like fallthrough,
fall-through, or fall-thru to use the LLVM_FALLTHROUGH macro instead.
llvm-svn: 278902
Summary: Thumb2 supports encoding immediates with specific patterns into mov.w by splatting the low 8 bits into other bytes.
I'm resubmitting this patch. The test case in the original commit
r277610 does not specify triple, so builds with differnt default triple
will have different output.
This patch fixed trile as thumb-darwin-apple.
Reviewers: john.brawn, jmolloy, bruno
Subscribers: jmolloy, aemerson, rengolin, samparker, llvm-commits
Differential Revision: https://reviews.llvm.org/D23090
llvm-svn: 277865
Summary: Thumb2 supports encoding immediates with specific patterns into mov.w by splatting the low 8 bits into other bytes.
Reviewers: john.brawn, jmolloy
Subscribers: jmolloy, aemerson, rengolin, samparker, llvm-commits
Differential Revision: https://reviews.llvm.org/D23090
llvm-svn: 277610
In this particular example we wouldn't want the smmls anyway (the value is
actually unused), but in general smmls does not provide the required flags
register so if that SUBE result is used we can't replace it.
llvm-svn: 277541
Retry r275776 (no changes, we suspect the issue was with another commit).
The current logic for handling inline asm operands in DAGToDAGISel interprets
the operands by looking for constants, which should represent the flags
describing the kind of operand we're dealing with (immediate, memory, register
def etc). The operands representing actual data are skipped only if they are
non-const, with the exception of immediate operands which are skipped explicitly
when a flag describing an immediate is found.
The oversight is that memory operands may be const too (e.g. for device drivers
reading a fixed address), so we should explicitly skip the operand following a
flag describing a memory operand. If we don't, we risk interpreting that
constant as a flag, which is definitely not intended.
Fixes PR26038
Differential Revision: https://reviews.llvm.org/D22103
llvm-svn: 276101
The current logic for handling inline asm operands in DAGToDAGISel interprets
the operands by looking for constants, which should represent the flags
describing the kind of operand we're dealing with (immediate, memory, register
def etc). The operands representing actual data are skipped only if they are
non-const, with the exception of immediate operands which are skipped explicitly
when a flag describing an immediate is found.
The oversight is that memory operands may be const too (e.g. for device drivers
reading a fixed address), so we should explicitly skip the operand following a
flag describing a memory operand. If we don't, we risk interpreting that
constant as a flag, which is definitely not intended.
Fixes PR26038
Differential Revision: https://reviews.llvm.org/D22103
llvm-svn: 275776
Thumb-1 doesn't have post-inc or pre-inc load or store instructions. However the LDM/STM instructions with writeback can function as post-inc load/store:
ldm r0!, {r1} @ load from r0 into r1 and increment r0 by 4
Obviously, this only works if the post increment is 4.
llvm-svn: 275540
This is a follow-up for r273544.
The end goal is to get rid of the isSwift / isCortexXY / isWhatever methods.
This commit also removes a command line flag that isn't used in any of the tests:
check-vmlx-hazards. It can be replaced easily with the mattr mechanism, since
this is now a subtarget feature.
There is still some work left regarding FeatureExpandMLx. In the past MLx
expansion was enabled for subtargets with hasVFP2(), until r129775 [1] switched
from that to isCortexA9, without too much justification.
In spite of that, the code performing MLx expansion still contains calls to
isSwift/isLikeA9, although the results of those are pretty clear given that
we're only enabling it for the A9.
We should try to enable it for all targets that have FeatureHasVMLxHazards, as
it seems to be closely related to that behaviour, and if that is possible try to
clean up the MLx expansion pass from all calls to isWhatever. This will require
some performance testing, so it will be done in another patch.
[1] http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20110418/119725.html
Differential Revision: http://reviews.llvm.org/D21798
llvm-svn: 274742
The important thing I was missing was ensuring newly added constants were kept in topological order. Repositioning the node is correct if the constant is newly added (so it has no topological ordering) but wrong if it already existed - positioning it next in the worklist would break the topological ordering.
Original commit message:
[Thumb] Select a BIC instead of AND if the immediate can be encoded more optimally negated
If an immediate is only used in an AND node, it is possible that the immediate can be more optimally materialized when negated. If this is the case, we can negate the immediate and use a BIC instead;
int i(int a) {
return a & 0xfffffeec;
}
Used to produce:
ldr r1, [CONSTPOOL]
ands r0, r1
CONSTPOOL: 0xfffffeec
And now produces:
movs r1, #255
adds r1, #20 ; Less costly immediate generation
bics r0, r1
llvm-svn: 274543
Fangrui Song