The VOP3 form of the V_CNDMASK_B32 instruction takes a carry-in
operand. The conversion to SDWA implies a conversion to VOP2 form
which reads from VCC instead.
Convert V_CNDMASK_B32_e64 instructions that might be converted to SDWA
to V_CNDMASK_B32_e32 first and introduce a copy of the carry-in operand
to VCC.
Closes#133431.
---------
Co-authored-by: Matt Arsenault <arsenm2@gmail.com>
There are V2S copies between vpgr16 and spgr32 in true16 mode. This is
caused by vgpr16 and sgpr32 both selectable by 16bit src in ISel.
When a V2S copy and its useMI are lowered to VALU, this patch check
1. If the generated new VALU is used by a true16 inst. Add subreg access
if necessary.
2. Legalize the V2S copy by replacing it to subreg_to_reg
an example MIR looks like:
```
%2:sgpr_32 = COPY %1:vgpr_16
%3:sgpr_32 = S_OR_B32 %2:sgpr_32, ...
%4:vgpr_16 = V_ADD_F16_t16 %3:sgpr_32, ...
```
currently lowered to
```
%2:vgpr_32 = COPY %1:vgpr_16
%3:vgpr_32 = V_OR_B32 %2:vgpr_32, ...
%4:vgpr_16 = V_ADD_F16_t16 %3:vgpr_32, ...
```
after this patch
```
%2:vgpr_32 = SUBREG_TO_REG 0, %1:vgpr_16, lo16
%3:vgpr_32 = V_OR_B32 %2:vgpr_32, ...
%4:vgpr_16 = V_ADD_F16_t16 %3.lo16:vgpr_32, ...
```
Added a srl pattern for true16 flow. Changing right shift 16bit to a
reg_sequence
`srl vgpr32, 16 -> reg_sequence (vgpr32.hi16, 0)`
and finally it's lowered to two COPY
`vdst.lo16 = COPY vsrc.hi16`
`vdst.hi16 = COPY 0`
The benefits of this transform is allowing the following pass to
optimize out these copy.
This PR updates the SGPR layout to a striped caller/callee-saved design,
similar
to the VGPR layout.
To ensure that s30-s31 (return address), s32 (stack pointer), s33 (frame
pointer), and s34 (base pointer) remain callee-saved, the striped layout
starts
from s40, with a stripe width of 8. The last stripe is 10 wide instead
of 8 to
avoid ending with a 2-wide stripe.
Fixes#113782.
Occupancy (i.e., the number of waves per EU) depends, in addition to
register usage, on per-workgroup LDS usage as well as on the range of
possible workgroup sizes. Mirroring the latter, occupancy should
therefore be expressed as a range since different group sizes generally
yield different achievable occupancies.
`getOccupancyWithLocalMemSize` currently returns a scalar occupancy
based on the maximum workgroup size and LDS usage. With respect to the
workgroup size range, this scalar can be the minimum, the maximum, or
neither of the two of the range of achievable occupancies. This commit
fixes the function by making it compute and return the range of
achievable occupancies w.r.t. workgroup size and LDS usage; it also
renames it to `getOccupancyWithWorkGroupSizes` since it is the range of
workgroup sizes that produces the range of achievable occupancies.
Computing the achievable occupancy range is surprisingly involved.
Minimum/maximum workgroup sizes do not necessarily yield maximum/minimum
occupancies i.e., sometimes workgroup sizes inside the range yield the
occupancy bounds. The implementation finds these sizes in constant time;
heavy documentation explains the rationale behind the sometimes
relatively obscure calculations.
As a justifying example, consider a target with 10 waves / EU, 4 EUs/CU,
64-wide waves. Also consider a function with no LDS usage and a flat
workgroup size range of [513,1024].
- A group of 513 items requires 9 waves per group. Only 4 groups made up
of 9 waves each can fit fully on a CU at any given time, for a total of
36 waves on the CU, or 9 per EU. However, filling as much as possible
the remaining 40-36=4 wave slots without decreasing the number of groups
reveals that a larger group of 640 items yields 40 waves on the CU, or
10 per EU.
- Similarly, a group of 1024 items requires 16 waves per group. Only 2
groups made up of 16 waves each can fit fully on a CU ay any given time,
for a total of 32 waves on the CU, or 8 per EU. However, removing as
many waves as possible from the groups without being able to fit another
equal-sized group on the CU reveals that a smaller group of 896 items
yields 28 waves on the CU, or 7 per EU.
Therefore the achievable occupancy range for this function is not [8,9]
as the group size bounds directly yield, but [7,10].
Naturally this change causes a lot of test churn as instruction
scheduling is driven by achievable occupancy estimates. In most unit
tests the flat workgroup size range is the default [1,1024] which,
ignoring potential LDS limitations, would previously produce a scalar
occupancy of 8 (derived from 1024) on a lot of targets, whereas we now
consider the maximum occupancy to be 10 in such cases. Most tests are
updated automatically and checked manually for sanity. I also manually
changed some non-automatically generated assertions when necessary.
Fixes#118220.
MSG_DEALLOC_VGPRS slows down very small waveslot limited kernels. It's
been identified this message is only really needed for VGPR limited
kernels. A kernel becomes VGPR limited if a total number of VGPRs per
SIMD / number of used VGPRs is more than a number of wave slots.
This reverts commit adaff46d087799072438dd744b038e6fd50a2d78.
Drop the -O3 checks from default-attributes.hip. I don't know why they
are different on some bots but reverting this is far too disruptive.
Removing it from the codegen pipeline induces a lot of test churn
because llc is no longer optimizing out implicit arguments to kernels.
Mostly mechanical, but there are some creative test updates. I preferred
to take the changes as-is in tests where the ABI isn't relevant. In
cases where it's more relevant, or the optimize out logic was too
ingrained in the test, I pre-run the optimization. Some cases manually
add attributes to disable inputs.
gfx11 chips may, in some conditions, behave incorrectly with S_CLAUSE
instructions (hard clauses) containing more than 32 operations (that is,
whose arguments exceed 0x1f). However, gfx10 targets will work
successfully with clauses of up to length 63.
Therefore, define the MaxHardClauseLength property on GCNSubtarget and
make it a subtarget feature via tablegen, thus allowing us to specify,
both now and in the future, the maximum viable size of clauses on
various hardware from the tablegen definition. If MaxHardClauseLength is
0, which is the default, the hardware does not support hard clauses.
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.
A bitcast of <10 x i32> to <5 x i64> was ending up on the
stack. Instead of doing that, handle the case where the new type
doesn't evenly divide but the elements do. Extract the individual
elements and pad with undef.
Avoids stack usage for bitcasts involving <5 x i64>. In some of these
cases, later optimizations actually eliminated the stack objects but
left behind the unused temporary stack object to final emission.
Fixes: SWDEV-377548
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
Use GCNHazardRecognizer in postra sched.
Updated tests for the new schedules.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109536
Change-Id: Ia86ba2ae168f12fb34b4d8efdab491f84d936cde
Treat a non-atomic volatile load and store as a relaxed atomic at
system scope for the address spaces accessed. This will ensure all
relevant caches will be bypassed.
A volatile atomic is not changed and still only bypasses caches upto
the level specified by the SyncScope operand.
Differential Revision: https://reviews.llvm.org/D94214
tryLatency compares two sched candidates. For the top zone it prefers
the one with lesser depth, but only if that depth is greater than the
total latency of the instructions we've already scheduled -- otherwise
its latency would be hidden and there would be no stall.
Unfortunately it only tests the depth of one of the candidates. This can
lead to situations where the TopDepthReduce heuristic does not kick in,
but a lower priority heuristic chooses the other candidate, whose depth
*is* greater than the already scheduled latency, which causes a stall.
The fix is to apply the heuristic if the depth of *either* candidate is
greater than the already scheduled latency.
All this also applies to the BotHeightReduce heuristic in the bottom
zone.
Differential Revision: https://reviews.llvm.org/D72392
We can produce such vectors in the Promote Alloca pass,
but we are unable to use movrel to operate it and lower
via scratch. Making it legal makes SI_INDIRECT patterns
work.
There is more work to do in subsequent changes:
1. We initialize m0 twice to access each dword. It shall
be possible to only do it once and increment base register
number instead.
2. We also need v16i64/v16f64 but these first need to be
added to tablegen.
Differential Revision: https://reviews.llvm.org/D79808
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
Summary:
- `dead-mi-elimination` assumes MIR in the SSA form and cannot be
arranged after phi elimination or DeSSA. It's enhanced to handle the
dead register definition by skipping use check on it. Once a register
def is `dead`, all its uses, if any, should be `undef`.
- Re-arrange the DIE in RA phase for AMDGPU by placing it directly after
`detect-dead-lanes`.
- Many relevant tests are refined due to different register assignment.
Reviewers: rampitec, qcolombet, sunfish
Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72709
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
The pre-RA scheduler does load/store clustering, but post-RA
scheduler undoes it. Add mutation to prevent it.
Differential Revision: https://reviews.llvm.org/D38014
llvm-svn: 313670
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
An encoding does not allow to use SDWA in an instruction with
scalar operands, either literals or SGPRs. That is however possible
to copy these operands into a VGPR first.
Several copies of the value are produced if multiple SDWA conversions
were done. To cleanup MachineLICM (to hoist copies out of loops),
MachineCSE (to remove duplicate copies) and SIFoldOperands (to replace
SGPR to VGPR copy with immediate copy right to the VGPR) runs are added
after the SDWA pass.
Differential Revision: https://reviews.llvm.org/D33583
llvm-svn: 304219
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
computeKnownBits didn't handle fp_to_fp16 to report
the high bits as 0. ARM maps the generic node to an instruction
that does not modify the high bits of the register, so introduce
a target node where the high bits are known 0.
llvm-svn: 297873
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
Canonicalize a select with a constant to the false side. This
enables more instruction shrinking opportunities since an
inline immediate can be used for the false side of v_cndmask_b32_e32.
This seems to usually be better but causes some code size regressions
in some tests.
llvm-svn: 290372
