PR #66334 tried to renumber slot indexes before register allocation, but
the numbering was still affected by list entries for instructions which
had been erased. Fix this to make the register allocator's live range
length heuristics even less dependent on the history of how instructions
have been added to and removed from SlotIndexes's maps.
RegAllocGreedy uses SlotIndexes::getApproxInstrDistance to approximate
the length of a live range for its heuristics. Renumbering all slot
indexes with the default instruction distance ensures that this estimate
will be as accurate as possible, and will not depend on the history of
how instructions have been added to and removed from SlotIndexes's maps.
This also means that enabling -early-live-intervals, which runs the
SlotIndexes analysis earlier, will not cause large amounts of churn due
to different register allocator decisions.
Occupancy is expressed as waves per SIMD. This means that we need to
take into account the number of SIMDs per "CU" or, to be more precise,
the number of SIMDs over which a workgroup may be distributed.
getOccupancyWithLocalMemSize was wrong because it didn't take SIMDs
into account at all.
At the same time, we need to take into account that WGP mode offers
access to a larger total amount of LDS, since this can affect how
non-power-of-two LDS allocations are rounded. To make this work
consistently, we distinguish between (available) local memory size and
addressable local memory size (which is always limited by 64kB on
gfx10+, even with WGP mode).
This change results in a massive amount of test churn. A lot of it is
caused by the fact that the default work group size is 1024, which means
that (due to rounding effects) the default occupancy on older hardware
is 8 instead of 10, which affects scheduling via register pressure
estimates. I've adjusted most tests by just running the UTC tools, but
in some cases I manually changed the work group size to 32 or 64 to make
sure that work group size chunkiness has no effect.
Differential Revision: https://reviews.llvm.org/D139468
Use shufflevector to do the subvector extracts. This allows a lot more
load merging on AMDGPU and also on NVPTX when <2 x half> is involved.
Differential Revision: https://reviews.llvm.org/D117219
Using a BufferSize of one for memory ProcResources will result in better
ILP since it more accurately models the dependencies between memory ops
and their consumers on an in-order processor. After this change, the
scheduler will treat the data edges from loads as blocking so that
stalls are guaranteed when waiting for data to be retreaved from memory.
Since we don't actually track waitcnt here, this should do a better job
at modeling their behavior.
Practically, this means that the scheduler will trigger the 'STALL'
heuristic more often.
This type of change needs to be evaluated experimentally. Preliminary
results are positive.
Fixes: SWDEV-282962
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D114777
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.
Use GCNHazardRecognizer in postra sched.
Updated tests for the new schedules.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109536
Change-Id: Ia86ba2ae168f12fb34b4d8efdab491f84d936cde
Clustering loads has caching benefits, but as far as I know there is no
advantage to clustering stores on any AMDGPU subtargets.
The disadvantage is that it tends to increase register pressure and
restricts scheduling freedom.
Differential Revision: https://reviews.llvm.org/D85530
If it is load cluster, we don't need to create the dependency edges(SUb->reg) from SUb to SUa
as they both depend on the base register "reg"
+-------+
+----> reg |
| +---+---+
| ^
| |
| |
| |
| +---+---+
| | SUa | Load 0(reg)
| +---+---+
| ^
| |
| |
| +---+---+
+----+ SUb | Load 4(reg)
+-------+
But if it is store cluster, we need to create it as follow shows to avoid the instruction store
depend on scheduled in-between SUb and SUa.
+-------+
+----> reg |
| +---+---+
| ^
| | Missing +-------+
| | +-------------------->+ y |
| | | +---+---+
| +---+-+-+ ^
| | SUa | Store x 0(reg) |
| +---+---+ |
| ^ |
| | +------------------------+
| | |
| +---+--++
+----+ SUb | Store y 4(reg)
+-------+
Reviewed By: evandro, arsenm, rampitec, foad, fhahn
Differential Revision: https://reviews.llvm.org/D72031
We are relying on atrificial DAG edges inserted by the
MemOpClusterMutation to keep loads and stores together in the
post-RA scheduler. This does not work all the time since it
allows to schedule a completely independent instruction in the
middle of the cluster.
Removed the DAG mutation and added pass to bundle already
clustered instructions. These bundles are unpacked before the
memory legalizer because it does not work with bundles but also
because it allows to insert waitcounts in the middle of a store
cluster.
Removing artificial edges also allows a more relaxed scheduling.
Differential Revision: https://reviews.llvm.org/D72737
Added support for dwordx3 for most load/store types, but not DS, and not
intrinsics yet.
SI (gfx6) does not have dwordx3 instructions, so they are not enabled
there.
Some of this patch is from Matt Arsenault, also of AMD.
Differential Revision: https://reviews.llvm.org/D58902
Change-Id: I913ef54f1433a7149da8d72f4af54dbb13436bd9
llvm-svn: 356659
See https://reviews.llvm.org/D47106 for details.
Reviewed By: probinson
Differential Revision: https://reviews.llvm.org/D47171
This commit drops that patch's changes to:
llvm/test/CodeGen/NVPTX/f16x2-instructions.ll
llvm/test/CodeGen/NVPTX/param-load-store.ll
For some reason, the dos line endings there prevent me from commiting
via the monorepo. A follow-up commit (not via the monorepo) will
finish the patch.
llvm-svn: 336843
This replaces most argument uses with loads, but for
now not all.
The code in SelectionDAG for calling convention lowering
is actively harmful for amdgpu_kernel. It attempts to
split the argument types into register legal types, which
results in low quality code for arbitary types. Since
all kernel arguments are passed in memory, we just want the
raw types.
I've tried a couple of methods of mitigating this in SelectionDAG,
but it's easier to just bypass this problem alltogether. It's
possible to hack around the problem in the initial lowering,
but the real problem is the DAG then expects to be able to use
CopyToReg/CopyFromReg for uses of the arguments outside the block.
Exposing the argument loads in the IR also has the advantage
that the LoadStoreVectorizer can merge them.
I'm not sure the best approach to dealing with the IR
argument list is. The patch as-is just leaves the IR arguments
in place, so all the existing code will still compute the same
kernarg size and pointlessly lowers the arguments.
Arguably the frontend should emit kernels with an empty argument
list in the first place. Alternatively a dummy array could be
inserted as a single argument just to reserve space.
This does have some disadvantages. Local pointer kernel arguments can
no longer have AssertZext placed on them as the equivalent !range
metadata is not valid on pointer typed loads. This is mostly bad
for SI which needs to know about the known bits in order to use the
DS instruction offset, so in this case this is not done.
More importantly, this skips noalias arguments since this pass
does not yet convert this to the equivalent !alias.scope and !noalias
metadata. Producing this metadata correctly seems to be tricky,
although this logically is the same as inlining into a function which
doesn't exist. Additionally, exposing these loads to the vectorizer
may result in degraded aliasing information if a pointer load is
merged with another argument load.
I'm also not entirely sure this is preserving the current clover
ABI, although I would greatly prefer if it would stop widening
arguments and match the HSA ABI. As-is I think it is extending
< 4-byte arguments to 4-bytes but doesn't align them to 4-bytes.
llvm-svn: 335650
This has two main components. First, widen
widen short constant loads in DAG when they have
the correct alignment. This is already done a bit in
AMDGPUCodeGenPrepare, since that has access to
DivergenceAnalysis. This can't help kernarg loads
created in the DAG. Start to use DAG divergence analysis
to help this case.
The second part is to avoid kernel argument lowering
breaking the alignment of short vector elements because
calling convention lowering wants to split everything
into legal register types.
When loading a split type, load the nearest 4-byte aligned
segment and shift to get the desired bits. This extra
load of the earlier argument piece ends up merging,
and the bit extract hopefully folds out.
There are a number of improvements and regressions with
this, but I think as-is this is a better compromise between
several of the worst parts of SelectionDAG.
Particularly when i16 is legal, this produces worse code
for i8 and i16 element vector kernel arguments. This is
partially due to the very weak load merging the DAG does.
It only looks for fairly specific combines between pairs
of loads which no longer appear. In particular this
causes v4i16 loads to be split into 2 components when
previously the two halves were merged.
Worse, because of the newly introduced shifts, there
is a lot more unnecessary vector packing and unpacking code
emitted. At least some of this is due to reporting
false for isTypeDesirableForOp for i16 as a workaround for
the lack of divergence information in the DAG. The cases
where this happens it doesn't actually matter, but the
relevant code in SimplifyDemandedBits doens't have the context
to know to ignore this.
The use of the scalar cache is probably more important
than the mess of mostly scalar instructions doing this packing
and unpacking. Future work can fix this, possibly by making better
use of the new DAG divergence information for controlling promotion
decisions, or adding another version of shift + trunc + shift
combines that doesn't only know about the used types.
llvm-svn: 334180
This was just emitting loads with the ABI alignment
for the raw type. The true alignment is often better,
especially when an illegal vector type was scalarized.
The better alignment allows using a scalar load
more often.
llvm-svn: 333558
This usually results in better code. Fixes using
inline asm with short2, and also fixes having a different
ABI for function parameters between VI and gfx9.
Partially cleans up the mess used for lowering of the d16
operations. Making v4f16 legal will help clean this up more,
but this requires additional work.
llvm-svn: 332953
Try to avoid mutually exclusive features. Don't use
a real default GPU, and use a fake "generic". The goal
is to make it easier to see which set of features are
incompatible between feature strings.
Most of the test changes are due to random scheduling changes
from not having a default fullspeed model.
llvm-svn: 310258
Currently the default C calling convention functions are treated
the same as compute kernels. Make this explicit so the default
calling convention can be changed to a non-kernel.
Converted with perl -pi -e 's/define void/define amdgpu_kernel void/'
on the relevant test directories (and undoing in one place that actually
wanted a non-kernel).
llvm-svn: 298444
This switches to the workaround that HSA defaults to
for the mesa path.
This should be applied to the 4.0 branch.
Patch by Vedran Miletić <vedran@miletic.net>
llvm-svn: 292982
Limit register coalescer by not allowing it to artificially increase
size of registers beyond dword. Such super-registers are in fact
register sequences and not distinct HW registers.
With more super-regs we would need to allocate adjacent registers
and constraint regalloc more than needed. Moreover, our super
registers are overlapping. For instance we have VGPR0_VGPR1_VGPR2,
VGPR1_VGPR2_VGPR3, VGPR2_VGPR3_VGPR4 etc, which complicates registers
allocation even more, resulting in excessive spilling.
Differential Revision: https://reviews.llvm.org/D28782
llvm-svn: 292413
If the literal is being folded into src0, it doesn't matter
if it's an SGPR because it's being replaced with the literal.
Also fixes initially selecting 32-bit versions of some instructions
which also confused commuting.
llvm-svn: 281117
There was a combine before to handle the simple copy case.
Split this into handling loads and stores separately.
We might want to change how this handles some of the vector
extloads, since this can result in large code size increases.
llvm-svn: 274394
Allocating larger register classes first should give better allocation
results (and more importantly for myself, make the lit tests more stable
with respect to scheduler changes).
Patch by Matthias Braun
llvm-svn: 270312
Summary:
This includes a hazard recognizer implementation to replace some of
the hazard handling we had during frame index elimination.
Reviewers: arsenm
Subscribers: qcolombet, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D18602
llvm-svn: 268143
Summary:
The goal is for each operand type to have its own parse function and
at the same time share common code for tracking state as different
instruction types share operand types (e.g. glc/glc_flat, etc).
Introduce parseAMDGPUOperand which can parse any optional operand.
DPP and Clamp/OMod have custom handling for now. Sam also suggested
to have class hierarchy for operand types instead of table. This
can be done in separate change.
Remove parseVOP3OptionalOps, parseDS*OptionalOps, parseFlatOptionalOps,
parseMubufOptionalOps, parseDPPOptionalOps.
Reduce number of definitions of AsmOperand's and MatchClasses' by using common base class.
Rename AsmMatcher/InstPrinter methods accordingly.
Print immediate type when printing parsed immediate operand.
Use 'off' if offset/index register is unused instead of skipping it to make it more readable (also agreed with SP3).
Update tests.
Reviewers: tstellarAMD, SamWot, artem.tamazov
Subscribers: qcolombet, arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D19584
llvm-svn: 268015
On AMDGPU where operations i64 operations are often bitcasted to v2i32
and back, this pattern shows up regularly where it breaks some
expected combines on i64, such as load width reducing.
This fixes some test failures in a future commit when i64 loads
are changed to promote.
llvm-svn: 262397
The one regression in the builtin tests is in the read2 test which now
(again) has many extra copies, but this should be solved once the pass
is replaced with a DAG combine.
llvm-svn: 253974
