We were letting G_ANYEXT with a vcc register bank through, which was
incorrect and would select to an invalid copy. Fix this up like G_ZEXT
and G_SEXT. Also drop old code to fixup the non-boolean case in
RegBankSelect. We now have to perform that expansion during selection,
so there's no benefit to doing it during RegBankSelect.
Fixes integers that don't evenly divide to i32 pieces. We should
probably extract some of the code in the legalizer to start handling
argument breakdowns. I'm dissatisfied with the argument lowering's
handling of vectors for example, and we should not be producing the
weird G_EXTRACTs we do now.
I used the implementation for floor instead of round. It also turns
out the OpenCL builtin library wasn't using the round builtin, but
implemented the expanded form.
Summary:
Truncating the result of a merge means that most likely we could have done without merge in the first place and just used the input merge inputs directly. This can be done in three cases:
1. If the truncation result is smaller than the merge source, we can use the source in the trunc directly
2. If the sizes are the same, we can replace the register or use a copy
3. If the truncation size is a multiple of the merge source size, we can build a smaller merge
This gets rid of most of the larger, hard-to-legalize merges.
Reviewers: qcolombet, aditya_nandakumar, aemerson, paquette, arsenm, Petar.Avramovic
Reviewed By: arsenm
Subscribers: sdardis, jvesely, wdng, nhaehnle, rovka, jrtc27, atanasyan, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D75915
Instead, emit a trap and a warning. We force inlining of this
situation, so any function where this happens should be dead as
indirect or external calls are not yet supported. This should avoid
erroring on dead code.
This was failng on any pre-assigned copy to the VCC bank.
This is something of a workaround for the default implementation in
getInstrMappingImpl, and how it treats copy-like operations in
general.
Copy-like operations are considered to only have one result register
bank, rather than separate banks for each source like a normal
instruction. To avoid potentially mishandling reg_sequence with
impossible operand combinations, the generic implementation errors on
impossible costs. If the bank was already assigned, is treated it
as-if it were an unsatisfiable REG_SEQUENCE mapping. We really don't
get any value from any of what getInstrMappingImpl tries to do for
copies, so just directly emit the simple mapping we really want.
If the loaded memory size was smaller than the result size, this would
produce out of bounds memory accesses. I'm wondering if we need a
distinct narrow memory legalize action type, since a case I care about
is decomposing a 4-byte unaligned access into 4 extending loads, which
would leave the original result register type. I'm currently awkwardly
using narrowScalar to handle unaligned accesses that need to be split.
This avoids regressions in a future patch. I'm confused by the use of
the gfx9 usage legacy_mad. Was this a pointless instruction rename, or
uses fmul_legacy handling? Why is regular mac avilable in that case?
In case the source value ends up in a VGPR, insert a readfirstlane to
avoid producing an illegal copy later. If it turns out to be
unnecessary, it can be folded out.
When expanding scalar packed operations, we should not introduce
illegal vector casts LegalizerHelper introduces. We're not in a
legalizer context, and there's no RegBankSelect apply or legalize
worklist.
We weren't considering the packed case correctly, and this was passing
through to the selector. The selector only checked the size, so this
would incorrectly compile to a single 32-bit scalar add.
As usual, the LegalizerHelper is somewhat awkward to use from
applyMappingImpl. I think this is the first place we've needed
multi-step legalization here though.
Most importantly, this fixes ret i8. Also make sure to handle
signext/zeroext for odd types > i32. Some of the corresponding
argument passing fixes also need to be handled.
Use a hack to only enable this for GlobalISel.
Technically this also works with SelectionDAG, but the divergence
selection isn't reliable enough and a few cases fail, but I have no
desire to spend time writing the manual expansion code for it. The DAG
actually does a better job since it catches using v_add_lshl_u32 in
the mixed SGPR/VGPR cases.
This reverts commit 5583c2f2fb.
The lldb bot failure was a test that was fragile and sensitive to irrelevant
changes in instruction ordering. Re-committing this as the test should have
been skipped for AArch64 now.
Differential Revision: https://reviews.llvm.org/D75555
Create a wider source vector, and unmerge with dead defs like the
legalizer. The legalization handling for G_EXTRACT is incomplete, and
it's preferrable to keep everything in 32-bit pieces.
We should probably start moving these functions into utils, since we
have a growing number of places that do almost the same thing.
Since this is still largely relying on the DAG argument type lowering
code, this has inherited the problem where i16 vectors have a
different ABI on targets with and without legal i16. Switch to using a
target with legal i16, so the i16 vector argument tests are more
useful.
This changes the localizer to attempt intra-block localizer of instructions
that have local uses. This is useful because sometimes the entry block itself
has many uses of constant-like instructions, which would benefit from shortening
live ranges. Previously if an inst had no non-local uses, we wouldn't add it to
the list of instructions to attempt further intra-block localization.
This gives a 0.7% geomean code size improvement on CTMark.
Differential Revision: https://reviews.llvm.org/D75555
Summary:
pickNodeBidirectional tried to compare the best top candidate and the
best bottom candidate by examining TopCand.Reason and BotCand.Reason.
This is unsound because, after calling pickNodeFromQueue, Cand.Reason
does not reflect the most important reason why Cand was chosen. Rather
it reflects the most recent reason why it beat some other potential
candidate, which could have been for some low priority tie breaker
reason.
I have seen this cause problems where TopCand is a good candidate, but
because TopCand.Reason is ORDER (which is very low priority) it is
repeatedly ignored in favour of a mediocre BotCand. This is not how
bidirectional scheduling is supposed to work.
To fix this I changed the code to always compare TopCand and BotCand
directly, like the generic implementation of pickNodeBidirectional does.
This removes some uncommented AMDGPU-specific logic; if this logic turns
out to be important then perhaps it could be moved into an override of
tryCandidate instead.
Graphics shader benchmarking on gfx10 shows a lot more positive than
negative effects from this change.
Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB
Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68338
The current set of custom combines are only really useful after
legalization, so move them there. There is a lot of overlap in the
boilerplate here, but I think we do want a pretty different set of
combines before and after legalize. I think we will want a lot of
overlap between the post-legalize and a post-regbankselect combiner.
Changes the handling of odd breakdowns, and avoids using
G_EXTRACT/G_INSERT. Pad with undef to a wider size, and unmerge. Also
avoid introducing instructions for the fully undef components.
The legalizer helper functions are unusably awkward to perform the 3-5
part legalization. This needs to be widened, scalarized, lowered, and
we should avoid creating vector extends and truncates. Manually do all
of this and expand.
I tried to use some of the new tablegen features to avoid creating
different operand list permutations, but I still don't see a way to
programmatically build a source pattern dag.
Also add GlobalISel tests, which now all import successfully.
Some of the fneg fold tests are incorrect, which need to be fixed in a
future commit
G_SHUFFLE_VECTOR is legal since it theoretically may help match op_sel
for VOP3P instructions. Expand it in some other way in case it doesn't
fold into the use instructions.
We should try the generated matchers before the manual selection. This
means the patterns are now handling the common cases, but the manual
selection code is not yet dead. It's still handling the non-s32/s64
cases (like v2s16 and v2s32). Currently tablegen doesn't have a nice
way to have a single pattern that covers multiple types.
We have patterns for s_pack* selection, but they assume the inputs are
a build_vector with 16-bit inputs, not a truncating build
vector. Since there's still outstanding work for how to handle
mismatched result and source element vector operations, and since I'm
trying a different packed vector strategy than SelectionDAG, just
manually select this for now.
There are few differences from the DAG handling. First, the DAG
handling uses a primitive selection pattern instead of custom
legalizing it. Because of this, this makes use of source modifiers
while the DAG does not.
Also instead of promoting f16, try to use the f16 log/exp. There's no
f16 fmul_legacy, so widen just for the multiply, although I'm not sure
that's the best solution.
This looked through copies to find the source modifiers, which may
have been SGPR->VGPR copies added to avoid potential constant bus
violations. Re-insert a copy to a VGPR if this happens.
Use common check prefix, although update_mir_test_checks makes this
unnecessarily annoying. Also make sure to have uses in case that ever
ends up mattering.
AMDGPUCodeGenPrepare expands this most of the time, but not always. We
will always at least need a fallback option here. This is the 3rd
implementation of the same expansion in the backend. Eventually I
would like to eliminate the IR expansion (and the DAG version
obviously).
Currently the new legalizer path produces a better result, since the
IR expansion results in extra operations which need to be combined
out. Notably, the IR expansion results in multiplies by 0.
Produce an unmerge to a narrower type and introduce a narrower shift
if needed. I wasn't sure if there was a better way to parameterize the
target's preferred shift type for the GICombineRule, so manually call
the combine helper.
While looking at the output on real sized programs, there is a lot of
extra SGPR spilling compared to the DAG path. This seems to largely be
from all constants being SGPRs in the entry block.
Try to handle arbitrary scalar BFEs by packing the operands. The DAG
gives up on non-constant arguments. We're still missing any constant
folding, so we end up with pretty ugly code most of the time. Also
handle the 64-bit scalar case, which the DAG doesn't try to do.
The division expansions in AMDGPUCodeGenPrepare can't be relied on for
correctness, since they punt to later optimization and possibly
legalization in some cases. We still need a way to be able to write
tests for the legalizer versions of the expansion. This is mostly for
GlobalISel, since the expected optimzations is expecting aren't
implemented.
The interaction with the flag to expand 64-bit division in the IR is
pretty confusing, but these flags have different purposes.
This is more or less directly ported from the AMDGPU custom lowering
for FP_TO_FP16. I made a few minor fixups (using G_UNMERGE_VALUES
instead of creating shift/trunc to extract the two halves, and zexting
an inverted compare instead of select_cc).
This also does not include the fast math expansion the DAG which
converts to f32 and then to f16. I think that belongs in a
pre-legalize combine instead.
This reverts commit 80a34ae311 with fixes.
Previously, since bots turning on EXPENSIVE_CHECKS are essentially turning on
MachineVerifierPass by default on X86 and the fact that
inline-asm-avx-v-constraint-32bit.ll and inline-asm-avx512vl-v-constraint-32bit.ll
are not expected to generate functioning machine code, this would go
down to `report_fatal_error` in MachineVerifierPass. Here passing
`-verify-machineinstrs=0` to make the intent explicit.
This reverts commit 80a34ae311 with fixes.
On bots llvm-clang-x86_64-expensive-checks-ubuntu and
llvm-clang-x86_64-expensive-checks-debian only,
llc returns 0 for these two tests unexpectedly. I tweaked the RUN line a little
bit in the hope that LIT is the culprit since this change is not in the
codepath these tests are testing.
llvm\test\CodeGen\X86\inline-asm-avx-v-constraint-32bit.ll
llvm\test\CodeGen\X86\inline-asm-avx512vl-v-constraint-32bit.ll
When SI_IF is inserted, it constrains the source register with a
register class, which was quite likely a G_ICMP. This was incorrectly
treating it as a scalar, and then applyMappingImpl would end up
producing invalid MIR since this was unexpected.
Also fix not using all VGPR sources for vcc outputs.
These are only legal on CI+. The test would fail in a debug build, but
not a release due to the partial selection since the pre-selection
legality assert only happens in a debug build.
Load extra bits if suitably aligned. This allows using widened
3-vector loads on SI, and fixes legalization for <9 x s32> (which LSV
apparently forms frequently on lowered kernel argument lists).
Fix incorrectly treating these as legal on SI. This should emit a
64-bit store and a 32-bit store.
I think all of the load and store rules are just about complete, but
due for a rewrite.
This reverts commit rGcd5b308b828e, rGcd5b308b828e, rG8cedf0e2994c.
There are issues to be investigated for polly bots and bots turning on
EXPENSIVE_CHECKS.
Vector indexing with a constant index should be folded out in the
legalizer, but this was accidentally falling through. This would
produce the indexing operation with $noreg. Handle this case as a
dynamic index just in case a bug like this happens again in the
future.
We were failing to find constants that were casted. I feel like the
artifact combiner should have folded the constant in the trunc before
the custom lowering, but that doesn't happen.
We were executing this in a waterfall loop as a placeholder, but this
should really be converted to a MUBUF load. Also execute in a
waterfall loop if the resource isn't an SGPR. This is a case where the
DAG handling was wrong because doing the right thing was too hard.
Currently, this will mishandle 96-bit loads. There's currently no way
to track the original memory size with an MMO, so these loads will be
widened andd the resulting memory size will be 128-bits.
The type passed to lower was invalid, so I'm not sure how this was
even working before. The source and destination type also do not have
to match, so make sure to use the right ones.
The legalizer produces a lot of these, and they make reading legalized
MIR annoying. For some reason, this does seem to sometimes introduce
copies of implicit def, which is dumb.
Really the intrinsic definition is wrong, but work around this
here. The DAG lowering introduces an MMO. We have to introduce a new
operation to avoid the verifier complaining about the missing mayLoad.
Use cmp ord instead of cmp_class compared to the DAG version for the
nan check, but mostly try to match the existsing pattern.
I think the sign doesn't matter for fract, so we could do a little
better with the source modifier matching.
I think this is also still broken as in D22898, but I'm leaving it
as-is for now while I don't have an SI system to test on.
Rewrite the result register pair into the expected sinigle register
format in the legalizer.
I'm also operating under the assumption that TFE doesn't apply to
stores or atomics, but don't know if this is true or not.
The 96-bit results need to be widened.
I find the interaction between LegalizerHelper and MIRBuilder somewhat
awkward. The custom legalization is called by the LegalizerHelper, but
then does not have access to the helper. You have to construct a new
helper, which then does not own the MachineIRBuilder, but does modify
it. Maybe custom legalization should be passed the helper?
If we have s_pack_* instructions, legalize this to
G_BUILD_VECTOR_TRUNC from s32 elements. This is closer to how how the
s_pack_* instructions really behave.
If we don't have s_pack_ instructions, expand this by creating a merge
to s32 and bitcasting. This expands to the expected bit operations. I
think this eventually should go in a new bitcast legalize action type
in LegalizerHelper.
We already directly emit the shift operations in RegBankSelect for the
vector case. This could possibly be cleaned up, but I also may want to
defer doing this expansion to selection anyway. I'll see about that
when I try to actually match VOP3P instructions.
This breaks the selection of the build_vector since tablegen doesn't
know how to match G_BUILD_VECTOR_TRUNC yet, so just xfail it for now.
Prepare to accurately track the future denormal-fp-math attribute
changes. The way to actually set these separately is not wired in yet.
This is just a mechanical change, and mostly still assumes the input
and output mode match. This should be refined for some cases. For
example, fcanonicalize lowering should use the flushing variant if
either input or output flushing is enabled
Start using a new strategy with a combination of merge and unmerges.
This allows scalarizing before lowering, which in cases like
<2 x s128> avoids producing giant illegal shifts.
This fixes legalizations of global stores > 128-bits. It seems work is
needed on how this split actually occurs. For example, we get the
right code for s160, with an s128 and s32 load, but get 5 s32 loads
for <5 x s32>.
Try out using combine definition rules.
This really should be a post-legalizer combine, but the combiner pass
is currently pre-legalize. Most of the target combines are really
post-legalize, so we should probably move the pass.
I believe this also fixes bugs with CI 32-bit handling, which was
incorrectly skipping offsets that look like signed 32-bit values. Also
validate the offsets are dword aligned before folding.
On targets that don't have the normal packed f16 layout, handle these
during legalization. Directly modify the register types. We can infer
this was a d16 load based on the mem operand size during selection.
A16 operands should possibly be handled here as well, but don't worry
about that yet.
This trivially avoids violating the constant bus restriction.
Previously this was allowing one SGPR in the first source
operand, which technically also avoided violating this for most
operations (but not for special cases reading vcc).
We do need to write some new, smarter operand folds to pick the
optimal SGPR to use in some kind of post-isel fold, but that's purely
an optimization.
I was originally thinking we would pick which operands should be SGPRs
in RegBankSelect, but I think this isn't really manageable. There
would be additional complexity to handle every G_* instruction, and
then any nontrivial instruction patterns would need to know when to
avoid violating it, which is likely to be very error prone.
I think having all inputs being canonically copies to VGPRs will
simplify the operand folding logic. The current folding we do is
backwards, and only considers one operand at a time, relative to
operands it already has. It therefore poorly handles the case where
there is already a constant bus operand user. If all operands are
copies, it's somewhat simpler to consider all input operands at once
to choose the optimal constant bus user.
Since the failure mode for constant bus violations is now a verifier
error and not an selection failure, this moves towards a place where
we can turn on the fallback mode. The SGPR copy folding optimizations
can be left for later.
For pow2 constants we should use G_SHL for pattern matching (and perf)
purposes later.
Vector support not yet implemented.
Differential Revision: https://reviews.llvm.org/D73659
Irritatingly the failure output is different in release vs. debug
because of the legality check is removed without asserts, so a register
ends up constrained only in release builds.
Fixes selection for scalar G_SMULH/G_UMULH. Also switches to using
tablegen selected add/sub, which switch to the signed version of the
opcode. This matches the current DAG behavior. We can't drop the
manual selection for add/sub yet, because it's still both for VALU
add/sub and for G_PTR_ADD.
Convert to the style most others use with one test instruction per
function, and use an implicit use to ensure the result register class
is constrained.
Change-Id: I6109148b0e3c80aa5535796a37abca583c19a936
This should be no problem to support with a pattern, but it turns out
there are just too many yaks to shave. The main problem is in the DAG
emitter, which I have no desire to sink effort into fixing.
If we had a bit to disable patterns in the DAG importer, fixing the
GlobalISelEmitter is more manageable.
Trivial type predicates should be moved into the tablegen pattern
itself, and not checked inside complex patterns. This eliminates a
redundant complex pattern, and fixes select source modifiers for
GlobalISel.
I have further patches which fully handle select in tablegen and
remove all of the C++ selection, although it requires the ugliness to
support the entire range of legal register types.
Use intermediate instructions, unlike with buffer stores. This is
necessary because of the need to have an internal way to distinguish
between signed and unsigned extloads. This introduces some duplication
and near duplication with the buffer store selection path. The store
handling should maybe be moved into legalization to match and
eliminate the duplication.
Try to keep simple v2s16 cases as-is. This will more naturally map to
how the VOP3P op_sel modifiers work compared to the expansion
involving bitcasts and bitshifts.
This could maybe try harder with wider source vector types, although
that could be handled with a pre-legalize combine.
For some reason the flat/global atomics end up in the generated
matcher table in a different order from SelectionDAG. Use
AddedComplexity to prefer checking for global atomics first.
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
The other 3-op patterns should also be theoretically handled, but
currently there's a bug in the inferred pattern complexity.
I'm not sure what the error handling strategy should be for potential
constant bus violations. I think the correct strategy is to never
produce mixed SGPR and VGPR operands in a typical VOP instruction,
which will trivially avoid them. However, it's possible to still have
hand written MIR (or erroneously transformed code) with these
operands. When these fold, the restriction will be violated. We
currently don't have any verifiers for reg bank legality. For now,
just ignore the restriction.
It might be worth triggering a DAG fallback on verifier error.
The pattern is also mishandled by the generated matcher, so workaround
this as in the DAG path.
The existing DAG tests aren't particularly targeted to just this one
intrinsic. These also end up differing in scheduling from SGPR->VGPR
operand constraint copies.
The waterfall utility function blindly inserts a phi for every def in
the loop. We don't need this one to be preserved for every
iteration. Saves an extra phi and copy inside the loop body.
The result and source vector are going to be tied, so these need to be
the same bank.
The inserted value also needs to be broken down based on the result
bank, not the inserted value itself.
Handle dynamic vector extracts that use an index that's an add of a
constant offset into moving the base subregister of the indexing
operation.
Force the add into the loop in regbankselect, which will be recognized
when selected.
The current implementation of skip insertion (SIInsertSkip) makes it a
mandatory pass required for correctness. Initially, the idea was to
have an optional pass. This patch inserts the s_cbranch_execz upfront
during SILowerControlFlow to skip over the sections of code when no
lanes are active. Later, SIRemoveShortExecBranches removes the skips
for short branches, unless there is a sideeffect and the skip branch is
really necessary.
This new pass will replace the handling of skip insertion in the
existing SIInsertSkip Pass.
Differential revision: https://reviews.llvm.org/D68092
Pointers of unrecognized address spaces shoudl be treated as
global-like pointers. Even if loads and stores of them aren't handled,
dumb operations that just operate on the bits should work.
This was assuming the narrow target was the source type. Respect the
requested type when these don't match by using intermediate
merges. This avoids producing very wide, illegal shift expansions.
This was dropping the invariant metadata on dead argument loads, so
they weren't deleted.
Atomics still need to be fixed the same way. Also, apparently store
was never preserving dereferencable which should also be fixed.
Matt Arsenault