Adds a pass to the ARM backend that takes a v4i32
gather and transforms it into a call to MVE's
masked gather intrinsics.
Differential Revision: https://reviews.llvm.org/D71743
Convert ARMCodeGenPrepare into a generic type promotion pass by:
- Removing the insertion of arm specific intrinsics to handle narrow
types as we weren't using this.
- Removing ARMSubtarget references.
- Now query a generic TLI object to know which types should be
promoted and what they should be promoted to.
- Move all codegen tests into Transforms folder and testing using opt
and not llc, which is how they should have been written in the
first place...
The pass searches up from icmp operands in an attempt to safely
promote types so we can avoid generating unnecessary unsigned extends
during DAG ISel.
Differential Revision: https://reviews.llvm.org/D69556
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
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
Some more refactoring, like registering the IT Block pass, less cryptic
variable names, and some simplification of loops.
Differential Revision: https://reviews.llvm.org/D63419
llvm-svn: 363666
Create the ARMBasicBlockUtils class for tracking and querying basic
blocks sizes so we can use them when generating low-overhead loops.
Differential Revision: https://reviews.llvm.org/D63265
llvm-svn: 363530
Initial commit of a new pass to create vector predication blocks, called VPT
blocks, that are supported by the Armv8.1-M MVE architecture.
This is a first naive implementation. I.e., for 2 consecutive predicated
instructions I1 and I2, for example, it will generate 2 VPT blocks:
VPST
I1
VPST
I2
A more optimal implementation would obviously put instructions in the same VPT
block when they are predicated on the same condition and when it is allowed to
do this:
VPTT
I1
I2
We will address this optimisation with follow up patches when the groundwork is
in. Creating VPT Blocks is very similar to IT Blocks, which is the reason I
added this to Thumb2ITBlocks.cpp. This allows reuse of the def use analysis
that we need for the more optimal implementation.
VPT blocks cannot be nested in IT blocks, and vice versa, and so these 2 passes
cannot interact with each other. Instructions allowed in VPT blocks must
be MVE instructions that are marked as VPT compatible.
Differential Revision: https://reviews.llvm.org/D63247
llvm-svn: 363370
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
Arm specific codegen prepare is implemented to perform type promotion
on icmp operands, which can enable the removal of uxtb and uxth
(unsigned extend) instructions. This is possible because performing
type promotion before ISel alleviates this duty from the DAG builder
which has to perform legalisation, but has a limited view on data
ranges.
The pass visits any instruction operand of an icmp and creates a
worklist to traverse the use-def tree to determine whether the values
can simply be promoted. Our concern is values in the registers
overflowing the narrow (i8, i16) data range, so instructions marked
with nuw can be promoted easily. For add and sub instructions, we are
able to use the parallel dsp instructions to operate on scalar data
types and avoid overflowing bits. Underflowing adds and subs are also
permitted when the result is only used by an unsigned icmp.
Differential Revision: https://reviews.llvm.org/D48832
llvm-svn: 337687
Armv6 introduced instructions to perform 32-bit SIMD operations. The purpose of
this pass is to do some straightforward IR pattern matching to create ACLE DSP
intrinsics, which map on these 32-bit SIMD operations.
Currently, only the SMLAD instruction gets recognised. This instruction
performs two multiplications with 16-bit operands, and stores the result in an
accumulator. We will follow this up with patches to recognise SMLAD in more
cases, and also to generate other DSP instructions (like e.g. SADD16).
Patch by: Sam Parker and Sjoerd Meijer
Differential Revision: https://reviews.llvm.org/D48128
llvm-svn: 335850
Emit and use the TableGen instruction selector for ARM. At the moment,
this allows us to remove the hand-written code for selecting G_SDIV and
G_UDIV.
Future commits will focus on increasing the code coverage for it and
removing more dead code from the current instruction selector.
llvm-svn: 301905
Declare the ARMInstructionSelector in an anonymous namespace, to make it
more in line with the other targets which were migrated to this in
r299637 in order to avoid TableGen'erated headers being included in
non-GlobalISel builds.
llvm-svn: 301632
functions so that the size computation is available not only in ConstantIslands
but in other passes as well.
Differential Revision: https://reviews.llvm.org/D22640
llvm-svn: 276399
Initializing them in LLVMInitializeARMTarget() makes them visible early
enough for "llc -run-pass usage".
This required the pass to be renamed from "arm-load-store-opt" to
"arm-ldst-opt", because there already exists an arm-load-store-opt
cl::opt switch which would now clash with the passname getting added as
a switch in opt. On the bright side the pass name now matches the
DEBUG_TYPE name. Renamed "arm-prera-load-store-opt" to
"arm-repra-ldst-opt" as well for consistency.
llvm-svn: 275661
In PIC mode we were previously computing global variable addresses (or GOT
entry addresses) by adding the PC, the PC-relative GOT displacement and
the GOT-relative symbol/GOT entry displacement. Because the latter two
displacements are fixed, we ended up performing one more addition than
necessary.
This change causes us to compute addresses using a single PC-relative
displacement, resulting in a shorter code sequence. This reduces code size
by about 4% in a recent build of Chromium for Android.
As a result of this change we no longer need to compute the GOT base address
in the ARM backend, which allows us to remove the Global Base Reg pass and
SDAG lowering for the GOT.
We also now no longer use the GOT when addressing a symbol which is known
to be defined in the same linkage unit. Specifically, the symbol must have
either hidden visibility or a strong definition in the current module in
order to not use the the GOT.
This is a change from the previous behaviour where we would use the GOT to
address externally visible symbols defined in the same module. I think the
only cases where this could matter are cases involving symbol interposition,
but we don't really support that well anyway.
Differential Revision: http://reviews.llvm.org/D13650
llvm-svn: 251322
The patch is generated using this command:
tools/clang/tools/extra/clang-tidy/tool/run-clang-tidy.py -fix \
-checks=-*,llvm-namespace-comment -header-filter='llvm/.*|clang/.*' \
llvm/lib/
Thanks to Eugene Kosov for the original patch!
llvm-svn: 240137
on a per-function basis.
Previously some of the passes were conditionally added to ARM's pass pipeline
based on the target machine's subtarget. This patch makes changes to add those
passes unconditionally and execute them conditonally based on the predicate
functor passed to the pass constructors. This enables running different sets of
passes for different functions in the module.
rdar://problem/20542263
Differential Revision: http://reviews.llvm.org/D8717
llvm-svn: 239325
Add header guards to files that were missing guards. Remove #endif comments
as they don't seem common in LLVM (we can easily add them back if we decide
they're useful)
Changes made by clang-tidy with minor tweaks.
llvm-svn: 215558
be deleted. This will be reapplied as soon as possible and before
the 3.6 branch date at any rate.
Approved by Jim Grosbach, Lang Hames, Rafael Espindola.
This reverts commits r215111, 215115, 215116, 215117, 215136.
llvm-svn: 215154
I am sure we will be finding bits and pieces of dead code for years to
come, but this is a good start.
Thanks to Lang Hames for making MCJIT a good replacement!
llvm-svn: 215111
Still only 32-bit ARM using it at this stage, but the promotion allows
direct testing via opt and is a reasonably self-contained patch on the
way to switching ARM64.
At this point, other targets should be able to make use of it without
too much difficulty if they want. (See ARM64 commit coming soon for an
example).
llvm-svn: 206485
The previous situation where ATOMIC_LOAD_WHATEVER nodes were expanded
at MachineInstr emission time had grown to be extremely large and
involved, to account for the subtly different code needed for the
various flavours (8/16/32/64 bit, cmpxchg/add/minmax).
Moving this transformation into the IR clears up the code
substantially, and makes future optimisations much easier:
1. an atomicrmw followed by using the *new* value can be more
efficient. As an IR pass, simple CSE could handle this
efficiently.
2. Making use of cmpxchg success/failure orderings only has to be done
in one (simpler) place.
3. The common "cmpxchg; did we store?" idiom can be exposed to
optimisation.
I intend to gradually improve this situation within the ARM backend
and make sure there are no hidden issues before moving the code out
into CodeGen to be shared with (at least ARM64/AArch64, though I think
PPC & Mips could benefit too).
llvm-svn: 205525
ARM specific optimiztion, finding places in ARM machine code where 2 dmbs
follow one another, and eliminating one of them.
Patch by Reinoud Elhorst.
llvm-svn: 205409
a TargetMachine to construct (and thus isn't always available), to an
analysis group that supports layered implementations much like
AliasAnalysis does. This is a pretty massive change, with a few parts
that I was unable to easily separate (sorry), so I'll walk through it.
The first step of this conversion was to make TargetTransformInfo an
analysis group, and to sink the nonce implementations in
ScalarTargetTransformInfo and VectorTargetTranformInfo into
a NoTargetTransformInfo pass. This allows other passes to add a hard
requirement on TTI, and assume they will always get at least on
implementation.
The TargetTransformInfo analysis group leverages the delegation chaining
trick that AliasAnalysis uses, where the base class for the analysis
group delegates to the previous analysis *pass*, allowing all but tho
NoFoo analysis passes to only implement the parts of the interfaces they
support. It also introduces a new trick where each pass in the group
retains a pointer to the top-most pass that has been initialized. This
allows passes to implement one API in terms of another API and benefit
when some other pass above them in the stack has more precise results
for the second API.
The second step of this conversion is to create a pass that implements
the TargetTransformInfo analysis using the target-independent
abstractions in the code generator. This replaces the
ScalarTargetTransformImpl and VectorTargetTransformImpl classes in
lib/Target with a single pass in lib/CodeGen called
BasicTargetTransformInfo. This class actually provides most of the TTI
functionality, basing it upon the TargetLowering abstraction and other
information in the target independent code generator.
The third step of the conversion adds support to all TargetMachines to
register custom analysis passes. This allows building those passes with
access to TargetLowering or other target-specific classes, and it also
allows each target to customize the set of analysis passes desired in
the pass manager. The baseline LLVMTargetMachine implements this
interface to add the BasicTTI pass to the pass manager, and all of the
tools that want to support target-aware TTI passes call this routine on
whatever target machine they end up with to add the appropriate passes.
The fourth step of the conversion created target-specific TTI analysis
passes for the X86 and ARM backends. These passes contain the custom
logic that was previously in their extensions of the
ScalarTargetTransformInfo and VectorTargetTransformInfo interfaces.
I separated them into their own file, as now all of the interface bits
are private and they just expose a function to create the pass itself.
Then I extended these target machines to set up a custom set of analysis
passes, first adding BasicTTI as a fallback, and then adding their
customized TTI implementations.
The fourth step required logic that was shared between the target
independent layer and the specific targets to move to a different
interface, as they no longer derive from each other. As a consequence,
a helper functions were added to TargetLowering representing the common
logic needed both in the target implementation and the codegen
implementation of the TTI pass. While technically this is the only
change that could have been committed separately, it would have been
a nightmare to extract.
The final step of the conversion was just to delete all the old
boilerplate. This got rid of the ScalarTargetTransformInfo and
VectorTargetTransformInfo classes, all of the support in all of the
targets for producing instances of them, and all of the support in the
tools for manually constructing a pass based around them.
Now that TTI is a relatively normal analysis group, two things become
straightforward. First, we can sink it into lib/Analysis which is a more
natural layer for it to live. Second, clients of this interface can
depend on it *always* being available which will simplify their code and
behavior. These (and other) simplifications will follow in subsequent
commits, this one is clearly big enough.
Finally, I'm very aware that much of the comments and documentation
needs to be updated. As soon as I had this working, and plausibly well
commented, I wanted to get it committed and in front of the build bots.
I'll be doing a few passes over documentation later if it sticks.
Commits to update DragonEgg and Clang will be made presently.
llvm-svn: 171681
current IR-level pass.
The old SjLj EH pass has some problems, especially with the new EH model. Most
significantly, it violates some of the new restrictions the new model has. For
instance, the 'dispatch' table wants to jump to the landing pad, but we cannot
allow that because only an invoke's unwind edge can jump to a landing pad. This
requires us to mangle the code something awful. In addition, we need to keep the
now dead landingpad instructions around instead of CSE'ing them because the
DWARF emitter uses that information (they are dead because no control flow edge
will execute them - the control flow edge from an invoke's unwind is superceded
by the edge coming from the dispatch).
Basically, this pass belongs not at the IR level where SSA is king, but at the
code-gen level, where we have more flexibility.
llvm-svn: 140646
and MCSubtargetInfo.
- Added methods to update subtarget features (used when targets automatically
detect subtarget features or switch modes).
- Teach X86Subtarget to update MCSubtargetInfo features bits since the
MCSubtargetInfo layer can be shared with other modules.
- These fixes .code 16 / .code 32 support since mode switch is updated in
MCSubtargetInfo so MC code emitter can do the right thing.
llvm-svn: 134884
difficult on current ARM implementations for a few reasons.
1. Even though a single vmla has latency that is one cycle shorter than a pair
of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause
additional pipeline stall. So it's frequently better to single codegen
vmul + vadd.
2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to
stall for 4 cycles. We need to schedule them apart.
3. A vmla followed vmla is a special case. Obvious issuing back to back RAW
vmla + vmla is very bad. But this isn't ideal either:
vmul
vadd
vmla
Instead, we want to expand the second vmla:
vmla
vmul
vadd
Even with the 4 cycle vmul stall, the second sequence is still 2 cycles
faster.
Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough
but it isn't the optimial solution. This patch attempts to make it possible to
use vmla / vmls in cases where it is profitable.
A. Add missing isel predicates which cause vmla to be codegen'ed.
B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to
compute a fmul and a fmla.
C. Add additional isel checks for vmla, avoid cases where vmla is feeding into
fp instructions (except for the #3 exceptional case).
D. Add ARM hazard recognizer to model the vmla / vmls hazards.
E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the
vmla / vmls will trigger one of the special hazards.
Work in progress, only A+B are enabled.
llvm-svn: 120960
Pierre-vh