There are multiple possible ways to represent the X - urem X, Y pattern. SCEV was not canonicalizing, and thus, depending on which you analyzed, you could get different results. The sub representation appears to produce strictly inferior results in practice, so I decided to canonicalize to the Y * X/Y version.
The motivation here is that runtime unroll produces the sub X - (and X, Y-1) pattern when Y is a power of two. SCEV is thus unable to recognize that an unrolled loop exits because we don't figure out that the new unrolled step evenly divides the trip count of the unrolled loop. After instcombine runs, we convert the the andn form which SCEV recognizes, so essentially, this is just fixing a nasty pass ordering dependency.
The ARM loop hardware interaction in the test diff is opague to me, but the comments in the review from others knowledge of the infrastructure appear to indicate these are improvements in loop recognition, not regressions.
Differential Revision: https://reviews.llvm.org/D114018
This is the 'or' sibling for the fold added with:
D113212
https://alive2.llvm.org/ce/z/tgnp7K
Note that neither of these transforms is poison-safe,
but it does not seem to matter at this level. We have
had the scalar version of D113212 for a long time, so
this is just making optimizer behavior consistent.
We do not have the scalar version of *this* fold,
however, so that is another follow-up.
(X s< 0) ? Y : 0 --> (X s>> BW-1) & Y
We canonicalize to the icmp+select form in IR, and we already have this fold
for scalar select in SDAG, so I think it's an oversight that we don't have
the fold for vectors. It seems neutral for AArch64 and saves some instructions
on x86.
Whether we should also have the sibling folds for the inverse condition or
all-ones true value may depend on target-specific factors such as whether
there's an "and-not" instruction.
Differential Revision: https://reviews.llvm.org/D113212
To constant fold bitwise logic ops where we've legalized constant build vectors to a different type (e.g. v2i64 -> v4i32), this patch adds a basic ability to peek through the bitcasts and perform the constant fold on the inner operands.
The MVE predicate v2i64 regressions will be addressed by future support for basic v2i64 type support.
One of the yak shaving fixes for D113192....
Differential Revision: https://reviews.llvm.org/D113202
Currently when tail predicating loops, vpt blocks need to be created
with the vctp predicate in case we need to revert to non-tail predicated
form. This has the unfortunate side effect of severely hampering post-ra
scheduling at times as the instructions are already stuck in vpt blocks,
not allowed to be independently ordered.
This patch addresses that by just moving the creation of VPT blocks
later in the pipeline, after post-ra scheduling has been performed. This
allows more optimal scheduling post-ra before the vpt blocks are
created, leading to more optimal tail predicated loops.
Differential Revision: https://reviews.llvm.org/D113094
LLVM has the habit of turning adds with no common bits set into ors,
which means we need to detect them and treat them like adds again in the
MVE gather/scatter lowering pass.
Differential Revision: https://reviews.llvm.org/D112922
This teaches the MVE gather scatter lowering pass that SHL is
essentially the same as Mul, where we are able to optimize the
induction of a gather/scatter address by pushing them out of loops.
https://alive2.llvm.org/ce/z/wG4VyT
Differential Revision: https://reviews.llvm.org/D112920
The maximal value of a half is 0x7bff, which is 65504 when converted to
an integer. This patch teaches that to computeConstantRange to compute a
constant range with the correct maximum value.
https://alive2.llvm.org/ce/z/BV_Spbhttps://alive2.llvm.org/ce/z/Nwuqvb
The maximum value for a float converted in the same way is 3.4e38, which
requires 129bits of data. I have not added that here as integer types so
larger are rare, compared to integers types larger than 17 bits require
for half floats.
The MVE tests change because instsimplify happens to be run as a part of
the backend, where it doesn't tend to for other backends.
Differential Revision: https://reviews.llvm.org/D112694
Currently when creating tail predicated loops, we need to validate that
all the live-outs of a loop will be equivalent with and without tail
predication, and if they are not we cannot legally create a
tail-predicated loop, leaving expensive vctp and vpst instructions in
the loop. These notably can include register-allocation instructions
like stack loads and stores, and copys lowered from COPYs to MVE_VORRs.
Instead of trying to prove this is valid late in the pipeline, this
patch introduces a MQPRCopy pseudo instruction that COPY is lowered to.
This can then either be converted to a MVE_VORR where possible, or to a
couple of VMOVD instructions if not. This way they do not behave
differently within and outside of tail-predications regions, and we can
know by construction that they are always valid. The idea is that we can
do the same with stack load and stores, converting them to VLDR/VSTR or
VLDM/VSTM where required to prove tail predication is always valid.
This does unfortunately mean inserting multiple VMOVD instructions,
instead of a single MVE_VORR, but my experiments show it to be an
improvement in general.
Differential Revision: https://reviews.llvm.org/D111048
This adds a few more unpredicated intrinsics to sink splats to, in order
to create more qr instruction variants. Notably this includes
saddsat/uaddsat but also some of the unpredicated mve intrinsics.
Differential Revision: https://reviews.llvm.org/D110333
If the old instructions mentioned a subreg that the new instructions do
not, remove the subrange for that subreg.
For example, in TwoAddressInstructionPass::eliminateRegSequence, if a
use operand in the REG_SEQUENCE has the undef flag then we don't
generate a copy for it so after the elimination there should be no live
interval at all for the corresponding subreg of the def.
This is a small step towards switching TwoAddressInstructionPass over
from LiveVariables to LiveIntervals. Currently this path is only tested
if you explicitly enable -early-live-intervals.
Differential Revision: https://reviews.llvm.org/D110542
As we have to split blocks, we may be left in an invalid loop state
after a WLS is reverted to a DLS. Instead remember the WLS that could
not be fixed and revert them after finishing processing all other loops.
Differential Revision: https://reviews.llvm.org/D110567
The ARMLowOverheadLoops pass recalculates VPT block masks when it
converts VCMP's inside VPT blocks into VPT's. The function to do so
doesn't seem to handle debug info though, leading to invalid block
creation or asserts at compile time. Make sure the function skips any
debug info between the MVE instructions it inspects.
Differential Revision: https://reviews.llvm.org/D110564
In TwoAddressInstructionPass::processTiedPairs, update subranges of the
live interval for RegB as well as the main range.
This is a small step towards switching TwoAddressInstructionPass over
from LiveVariables to LiveIntervals. Currently this path is only tested
if you explicitly enable -early-live-intervals.
Differential Revision: https://reviews.llvm.org/D110526
We align non-fallthrough branches under Cortex-M at O3 to lead to fewer
instruction fetches. This improves that for the block after a LE or
LETP. These blocks will still have terminating branches until the
LowOverheadLoops pass is run (as they are not handled by analyzeBranch,
the branch is not removed until later), so canFallThrough will return
false. These extra branches will eventually be removed, leaving a
fallthrough, so treat them as such and don't add unnecessary alignments.
Differential Revision: https://reviews.llvm.org/D107810
Given:
- A jump table
- Which jumps to the next block
- The next block ends in a WLS
- Where the WLS conditionally jumps to block earlier in the program.
The Arm block placement pass would attempt to move the block containing
the WLS earlier, as the WLS instruction can only branch forward. In
doing so it would add a branch from the jumptable block to the WLS
block, thinking it previously fell-through.
This in itself would be fine, if a little inefficient, but the constant
island pass expects all instructions after a jump-table branch to have
been removed by analyzeBranch. So it gets confused and can assign the
same labels to multiple jump table blocks.
I've changed the condition to the same as used in analyzeBranch.
In repairIntervalsInRange, if the new instructions refer to subregs but
the old instructions did not, make sure any existing live interval for
the superreg is updated to have subranges. Also skip repairing any range
that we have recalculated from scratch, partly for efficiency but also
to avoids some cases that repairOldRegInRange can't handle.
The existing test/CodeGen/AMDGPU/twoaddr-regsequence.mir provides some
test coverage for this change: when TwoAddressInstructionPass converts
REG_SEQUENCE into subreg copies, the live intervals will now get
subranges and MachineVerifier will verify that the subranges are
correct. Unfortunately MachineVerifier does not complain if the
subranges are not present, so the test also passed before this patch.
This patch also fixes ~800 of the ~1500 failures in the whole CodeGen
lit test suite when -early-live-intervals is forced on.
Differential Revision: https://reviews.llvm.org/D110328
This allows VMOVL in tail predicated loops so long as the the vector
size the VMOVL is extending into is less than or equal to the size of
the VCTP in the tail predicated loop. These cases represent a
sign-extend-inreg (or zero-extend-inreg), which needn't block tail
predication as in https://godbolt.org/z/hdTsEbx8Y.
For this a vecsize has been added to the TSFlag bits of MVE
instructions, which stores the size of the elements that the MVE
instruction operates on. In the case of multiple size (such as a
MVE_VMOVLs8bh that extends from i8 to i16, the largest size was be
chosen). The sizes are encoded as 00 = i8, 01 = i16, 10 = i32 and 11 =
i64, which often (but not always) comes from the instruction encoding
directly. A unit test was added, and although only a subset of the
vecsizes are currently used, the rest should be useful for other cases.
Differential Revision: https://reviews.llvm.org/D109706
The vectorizer can sometimes make reverse shuffles from indices that
count down. In MVE, we don't have a 128bit rev instruction, but we can
select this to a VREV64 with some lane movs to swap the two halfs.
Ideally this would use VMOVD's, but only gets as far as VMOVS's at the
moment.
Differential Revision: https://reviews.llvm.org/D69510
This simple heuristic uses the estimated live range length combined
with the number of registers in the class to switch which heuristic to
use. This was taking the raw number of registers in the class, even
though not all of them may be available. AMDGPU heavily relies on
dynamically reserved numbers of registers based on user attributes to
satisfy occupancy constraints, so the raw number is highly misleading.
There are still a few problems here. In the original testcase that
made me notice this, the live range size is incorrect after the
scheduler rearranges instructions, since the instructions don't have
the original InstrDist offsets. Additionally, I think it would be more
appropriate to use the number of disjointly allocatable registers in
the class. For the AMDGPU register tuples, there are a large number of
registers in each tuple class, but only a small fraction can actually
be allocated at the same time since they all overlap with each
other. It seems we do not have a query that corresponds to the number
of independently allocatable registers. Relatedly, I'm still debugging
some allocation failures where overlapping tuples seem to not be
handled correctly.
The test changes are mostly noise. There are a handful of x86 tests
that look like regressions with an additional spill, and a handful
that now avoid a spill. The worst looking regression is likely
test/Thumb2/mve-vld4.ll which introduces a few additional
spills. test/CodeGen/AMDGPU/soft-clause-exceeds-register-budget.ll
shows a massive improvement by completely eliminating a large number
of spills inside a loop.
Currently, opaque pointers are supported in two forms: The
-force-opaque-pointers mode, where all pointers are opaque and
typed pointers do not exist. And as a simple ptr type that can
coexist with typed pointers.
This patch removes support for the mixed mode. You either get
typed pointers, or you get opaque pointers, but not both. In the
(current) default mode, using ptr is forbidden. In -opaque-pointers
mode, all pointers are opaque.
The motivation here is that the mixed mode introduces additional
issues that don't exist in fully opaque mode. D105155 is an example
of a design problem. Looking at D109259, it would probably need
additional work to support mixed mode (e.g. to generate GEPs for
typed base but opaque result). Mixed mode will also end up
inserting many casts between i8* and ptr, which would require
significant additional work to consistently avoid.
I don't think the mixed mode is particularly valuable, as it
doesn't align with our end goal. The only thing I've found it to
be moderately useful for is adding some opaque pointer tests in
between typed pointer tests, but I think we can live without that.
Differential Revision: https://reviews.llvm.org/D109290
I can't seem to wrap my head around the proper fix here,
we should be fine without this requirement, iff we can form this form,
but the naive attempt (https://reviews.llvm.org/D106317) has failed.
So just to unblock the release, put up a restriction.
Fixes https://bugs.llvm.org/show_bug.cgi?id=51125
As an extension to D107866, this adds store(fptosisat(..)) patterns,
similar to the existing fptosi patterns, to prevent unnecessarily moving
into gpr regs where we can use fp stores directly.
Differential Revision: https://reviews.llvm.org/D108378
This extends D107865 to the VFP insructions, lowering llvm.fptosi.sat
and llvm.fptoui.sat to VCVT instructions that inherently perform the
saturate.
Differential Revision: https://reviews.llvm.org/D107866
The semantics of tail predication loops means that the value of LR as an
instruction is executed determines the predicate. In other words:
mov r3, #3
DLSTP lr, r3 // Start tail predication, lr==3
VADD.s32 q0, q1, q2 // Lanes 0,1 and 2 are updated in q0.
mov lr, #1
VADD.s32 q0, q1, q2 // Only first lane is updated.
This means that the value of lr cannot be spilled and re-used in tail
predication regions without potentially altering the behaviour of the
program. More lanes than required could be stored, for example, and in
the case of a gather those lanes might not have been setup, leading to
alignment exceptions.
This patch adds a new lr predicate operand to MVE instructions in order
to keep a reference to the lr that they use as a tail predicate. It will
usually hold the zeroreg meaning not predicated, being set to the LR phi
value in the MVETPAndVPTOptimisationsPass. This will prevent it from
being spilled anywhere that it needs to be used.
A lot of tests needed updating.
Differential Revision: https://reviews.llvm.org/D107638
This adds lowering of the llvm.fptosi.sat and llvm.fptoui.sat intinsics,
selecting a VCVT instruction which under MVE will inherently perform the
saturate.
Differential Revision: https://reviews.llvm.org/D107865
This adds extra MVE vector fptosi.sat and fptoui.sat tests, along with
adding or adjusting the existing scalar tests to cover more
architectures and instruction combinations.
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
Add a variant of mve-vqdmulh tests that uses min/max intrinsics
directly, including a scalar test that shows it misbehaving for min
intrinsics and a fix for the combine to prevent it from misbehaving.
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
