This should now handle everything except structs passed as multiple
registers.
I think most of the packing logic should be handled by
handleAssignments, but I'm unclear on what the contract is for
multiple registers. This is copying how x86 handles this.
This does change the behavior of the test_sgpr_alignment0 amdgpu_vs
test. I don't think shader arguments should try to follow the
alignment, and registers need to be repacked. I also don't think it
matters, since I think the pointers are packed to the beginning of the
argument list anyway.
llvm-svn: 366582
This is the more natural lowering, and presents more opportunities to
reduce 64-bit ops to 32-bit.
This should also help avoid issues graphics shaders have had with
64-bit values, and simplify argument lowering in globalisel.
llvm-svn: 366578
Extract the sources to the GCD of the original size and target size,
padding with implicit_def as necessary.
Also fix the case where the requested source type is wider than the
original result type. This was ignoring the type, and just using the
destination. Do the operation in the requested type and truncate back.
llvm-svn: 366367
Use an anyext to the requested type for the leftover operand to
produce a slightly wider type, and then truncate the final merge.
I have another implementation almost ready which handles arbitrary
widens, but I think it produces worse code in this example (which I
think is 90% due to not folding redundant copies or folding out
implicit_def users), so I wanted to add this as a baseline first.
llvm-svn: 366366
I think this manages to not break the DAG handling with the divergent
predicates because the stadalone divergent patterns end up with a
higher priority than the pattern on the instruction definition.
The 16-bit versions don't work yet.
llvm-svn: 366254
Now that the patterns use the new PatFrag address space support, the
only blocker to importing most load patterns is the addressing mode
complex patterns.
llvm-svn: 366237
Apparently the check for legal instructions during instruction
select does not happen without an asserts build, so these would
successfully select in release, and fail in debug.
Make s16 and/or/xor legal. These can just be selected directly
to the 32-bit operation, as is already done in SelectionDAG, so just
make them legal.
llvm-svn: 366210
If a 1-bit value is in a 32-bit VGPR, the scalar opcodes set SCC to
whether the result is 0. If the inputs are SCC, these can be copied to
a 32-bit SGPR to produce an SCC result.
llvm-svn: 366125
This is a hack until I come up with a better way of dealing with the
pseudo-register banks used for boolean values. If the use instruction
constrains the register, the selector for the def instruction won't
see that the bank was VCC. A 1-bit SReg_32 is could ambiguously have
been SCCRegBank or VCCRegBank in wave32.
This is necessary to successfully select branches with and and/or/xor
condition.
llvm-svn: 366120
The extra test change is correct, although how it arrives there is a
bug that needs work. With wave32, the test for isVCC ambiguously
reports true for an SCC or VCC source. A new allocatable pseudo
register class for SCC may be necesssary.
llvm-svn: 366119
Before 2018, mesa used to use byval interchangably with inreg, which
didn't really make sense. Fix tests still using it to avoid breaking
in a future commit.
llvm-svn: 365953
In SelectionDAG AMDGPU treated these as legal, but this was mostly
because the bitcasts required for FP types were painful. Theoretically
the bitpattern should eventually match to bfi, so don't bother trying
to get the patterns to import.
llvm-svn: 365583
Mostsly these would fail due to trying to use SI with a flat
operation. Implementing global loads with MUBUF is more work than
flat, so these won't be handled in the initial load selection.
Others fail because store of s64 won't initially work, as the current
set of patterns expect everything to be turned into v2i32.
llvm-svn: 365493
There are two main issues preventing us from generating immediate form shifts:
1) We have partial SelectionDAG imported support for G_ASHR and G_LSHR shift
immediate forms, but they currently don't work because the amount type is
expected to be an s64 constant, but we only legalize them to have homogenous
types.
To deal with this, first we introduce a custom legalizer to *only* custom legalize
s32 shifts which have a constant operand into a s64.
There is also an additional artifact combiner to fold zexts(g_constant) to a
larger G_CONSTANT if it's legal, a counterpart to the anyext version committed
in an earlier patch.
2) For G_SHL the importer can't cope with the pattern. For this I introduced an
early selection phase in the arm64 selector to select these forms manually
before the tablegen selector pessimizes it to a register-register variant.
Differential Revision: https://reviews.llvm.org/D63910
llvm-svn: 364994
The register bank for the destination of the sample argument copy was
wrong. We shouldn't be constraining each source to the result register
bank. Allow constraining the original register to the right size.
llvm-svn: 364928
If the requested source type an be used as a merge source type, create
a merge of merges. This avoids creating large, illegal extensions and
bit-ops directly to the result type.
llvm-svn: 364841
Tests don't cover the masked input path since non-kernel arguments
aren't lowered yet.
Test is copied directly from the existing test, with 2 additions.
llvm-svn: 364833
Replace the brcond for the 2 cases that act as branches. For now
follow how the current system works, although I think we can
eventually get rid of the pseudos.
llvm-svn: 364832
This needs to be extended to s32, and expanded into cmp+select. This
is relying on the fact that widenScalar happens to leave the
instruction in place, but this isn't a guaranteed property of
LegalizerHelper.
llvm-svn: 364831
The condition register bank must be scc or vcc so that a copy will be
inserted, which will be lowered to a compare.
Currently greedy unnecessarily forces using a VCC select.
llvm-svn: 364825
Also works around tablegen defect in selecting add with unused carry,
but if we have to manually select GEP, might as well handle add
manually.
llvm-svn: 364806
There are several things broken, but at least emit the right thing for
gfx9.
The import of the pattern with the unused carry out seems to not
work. Needs a special class for clamp, because OperandWithDefaultOps
doesn't really work.
llvm-svn: 364804
This was checking the size of the register with the value of the size,
which happens to be exec. Also fix assuming VCC is 64-bit to fix
wave32.
Also remove some untested handling for physical registers which is
skipped. This doesn't insert the V_CNDMASK_B32 if SCC is the physical
copy source. I'm not sure if this should be trying to handle this
special case instead of dealing with this in copyPhysReg.
llvm-svn: 364761
Change the interface of CallLowering::lowerFormalArguments to accept
several virtual registers for each formal argument, instead of just one.
This is a follow-up to D46018.
CallLowering::lowerReturn was similarly refactored in D49660. lowerCall
will be refactored in the same way in follow-up patches.
With this change, we forward the virtual registers generated for
aggregates to CallLowering. Therefore, the target can decide itself
whether it wants to handle them as separate pieces or use one big
register. We also copy the pack/unpackRegs helpers to CallLowering to
facilitate this.
ARM and AArch64 have been updated to use the passed in virtual registers
directly, which means we no longer need to generate so many
merge/extract instructions.
AArch64 seems to have had a bug when lowering e.g. [1 x i8*], which was
put into a s64 instead of a p0. Added a test-case which illustrates the
problem more clearly (it crashes without this patch) and fixed the
existing test-case to expect p0.
AMDGPU has been updated to unpack into the virtual registers for
kernels. I think the other code paths fall back for aggregates, so this
should be NFC.
Mips doesn't support aggregates yet, so it's also NFC.
x86 seems to have code for dealing with aggregates, but I couldn't find
the tests for it, so I just added a fallback to DAGISel if we get more
than one virtual register for an argument.
Differential Revision: https://reviews.llvm.org/D63549
llvm-svn: 364510
Scalar extends to s64 can use S_BFE_{I64|U64}, but vector extends need
to extend to the 32-bit half, and then to 64.
I'm not sure what the line should be between what RegBankSelect
handles, and what instruction select does, but for now I'm erring on
the side of RegBankSelect for future post-RBS combines.
llvm-svn: 364212
This needs different handling if the source is known to be a valid
condition or not. Handle turning it into shifts or a select during
regbankselect.
llvm-svn: 364186
We sometimes get poor code size because constants of types < 32b are legalized
as 32 bit G_CONSTANTs with a truncate to fit. This works but means that the
localizer can no longer sink them (although it's possible to extend it to do so).
On AArch64 however s8 and s16 constants can be selected in the same way as s32
constants, with a mov pseudo into a W register. If we make s8 and s16 constants
legal then we can avoid unnecessary truncates, they can be CSE'd, and the
localizer can sink them as normal.
There is a caveat: if the user of a smaller constant has to widen the sources,
we end up with an anyext of the smaller typed G_CONSTANT. This can cause
regressions because of the additional extend and missed pattern matching. To
remedy this, there's a new artifact combiner to generate the wider G_CONSTANT
if it's legal for the target.
Differential Revision: https://reviews.llvm.org/D63587
llvm-svn: 364075
This was ignoring the flag on fneg, and using the source instruction's
flags. Also fixes tests missing from r358702.
Note the expansion itself isn't correct without nnan, but that should
be fixed separately.
llvm-svn: 363637
A target intrinsic may be defined as possibly reading memory, but the
call site may have additional knowledge that it doesn't read
memory. The intrinsic lowering will expect the pessimistic assumption
of the intrinsic definition, so the chain should still be used.
I fixed the same bug in SelectionDAG in r287593.
llvm-svn: 363580
I keep using the wrong instruction when manually writing tests. This
really needs to check the number of operands, but I don't see an easy
way to do that right now.
llvm-svn: 363579
Avoid producing illegal register bank copies for reg_sequence and
phi. The default implementation assumes it is possible to pick any
operand's bank and use that for the result, introducing a copy for
operands with a different bank. This does not check for illegal
copies. It is not legal to introduce a VGPR->SGPR copy, so any VGPR
operand requires the result to be a VGPR.
The changes in getInstrMappingImpl aren't strictly necessary, since
AMDGPU now just bypasses this for reg_sequence/phi. This could be
replaced with an assert in case other targets run into this. It is
currently responsible for producing the error for unsatisfiable
copies, but this will be better served with a verifier check.
For phis, for now assume any undetermined operands must be
VGPRs. Eventually, this needs to be able to defer mapping these
operations. This also does not yet have a way to check for whether the
block is in a divergent region.
llvm-svn: 363410
This is ported from the custom AMDGPU DAG implementation. I think this
is a better default expansion than what the DAG currently uses, at
least if the target has CTLZ.
This implements the signed version in terms of the unsigned
conversion, which is implemented with bit operations. SelectionDAG has
several other implementations that should eventually be ported
depending on what instructions are legal.
llvm-svn: 361081
Other opcodes shouldn't be CSE'd until we can be sure debug info quality won't
be degraded.
This change also improves the IRTranslator so that in most places, but not all,
it creates constants using the MIRBuilder directly instead of first creating a
new destination vreg and then creating a constant. By doing this, the
buildConstant() method can just return the vreg of an existing G_CONSTANT
instead of having to create a COPY from it.
I measured a 0.2% improvement in compile time and a 0.9% improvement in code
size at -O0 ARM64.
Compile time:
Program base cse diff
test-suite...ark/tramp3d-v4/tramp3d-v4.test 9.04 9.12 0.8%
test-suite...Mark/mafft/pairlocalalign.test 2.68 2.66 -0.7%
test-suite...-typeset/consumer-typeset.test 5.53 5.51 -0.4%
test-suite :: CTMark/lencod/lencod.test 5.30 5.28 -0.3%
test-suite :: CTMark/Bullet/bullet.test 25.82 25.76 -0.2%
test-suite...:: CTMark/ClamAV/clamscan.test 6.92 6.90 -0.2%
test-suite...TMark/7zip/7zip-benchmark.test 34.24 34.17 -0.2%
test-suite :: CTMark/SPASS/SPASS.test 6.25 6.24 -0.1%
test-suite...:: CTMark/sqlite3/sqlite3.test 1.66 1.66 -0.1%
test-suite :: CTMark/kimwitu++/kc.test 13.61 13.60 -0.0%
Geomean difference -0.2%
Code size:
Program base cse diff
test-suite...-typeset/consumer-typeset.test 1315632 1266480 -3.7%
test-suite...:: CTMark/ClamAV/clamscan.test 1313892 1297508 -1.2%
test-suite :: CTMark/lencod/lencod.test 1439504 1423112 -1.1%
test-suite...TMark/7zip/7zip-benchmark.test 2936980 2904172 -1.1%
test-suite :: CTMark/Bullet/bullet.test 3478276 3445460 -0.9%
test-suite...ark/tramp3d-v4/tramp3d-v4.test 8082868 8033492 -0.6%
test-suite :: CTMark/kimwitu++/kc.test 3870380 3853972 -0.4%
test-suite :: CTMark/SPASS/SPASS.test 1434904 1434896 -0.0%
test-suite...Mark/mafft/pairlocalalign.test 764528 764528 0.0%
test-suite...:: CTMark/sqlite3/sqlite3.test 782092 782092 0.0%
Geomean difference -0.9%
Differential Revision: https://reviews.llvm.org/D60580
llvm-svn: 358369
The register index can only really be an SGPR. Lie that a VGPR index
is legal, and then rewrite the instruction in a waterfall loop to
handle the index.
llvm-svn: 357235
The artifact combiners push instructions which have been marked for deletion
onto an list for the legalizer to deal with on return. However, for trunc(ext)
combines the combiner routine recursively calls itself. When it does this the
dead instructions list may not be empty, and the other combiners don't expect
to be dealing with essentially invalid MIR (multiple vreg defs etc).
This change fixes it by ensuring that the dead instructions are processed on
entry into tryCombineInstruction.
As a result, this fix exposed a few places in tests where G_TRUNC instructions
were not being deleted even though they were dead.
Differential Revision: https://reviews.llvm.org/D59892
llvm-svn: 357101
The AMDGPU implementation of getReservedRegs depends on
MachineFunctionInfo fields that are parsed from the YAML section. This
was reserving the wrong register since it was setting the reserved
regs before parsing the correct one.
Some tests were relying on the default reserved set for the assumed
default calling convention.
llvm-svn: 357083
The AArch64 test was broken since the result register already had a
set register class, so this test was a no-op. The mapping verify call
would fail because the result size is not the same as the inputs like
in a copy or phi.
The AMDGPU testcases are half broken and introduce illegal VGPR->SGPR
copies which need much more work to handle correctly (same for phis),
but add them as a baseline.
llvm-svn: 356713
The constantness shouldn't change the register bank choice. We also
don't need to restrict this to only indexing VGPRs, since it's
possible to index SGPRs (but SelectionDAG made using this
difficult). Allow directly indexing SGPRs when appropriate.
llvm-svn: 356611
This is consistent with what SelectionDAG does and is much easier to
work with than the extract sequence with an artificial wide register.
For the AMDGPU control flow intrinsics, this was producing an s128 for
the i64, i1 tuple return. Any legalization that should apply to a real
s128 value would badly obscure the direct values that need to be seen.
llvm-svn: 356147
Narrow Scalar G_MUL for MIPS32.
Revisit NarrowScalar implementation in LegalizerHelper.
Introduce new helper function multiplyRegisters.
It performs generic multiplication of values held in multiple registers.
Generated instructions use only types NarrowTy and i1.
Destination can be same or two times size of the source.
Differential Revision: https://reviews.llvm.org/D58824
llvm-svn: 355814
Try to use concat_vectors. Also remove unnecessary assert on
pointers. Fixes asserting for <4 x s16> operations and 64-bit pointers
for AMDGPU.
llvm-svn: 354828
This is basically a pointer typed add, so shouldn't be any different.
This was assuming everything was an SGPR, which is not true.
Also cleanup legality for GEP. I don't seem to be seeing the problem
the hack marking s64 as a legal pointer type the comment mentions.
llvm-svn: 354067
clampScalar doesn't do anything for non-power-of-2 in range.
There should probably be a combination rule to reduce the number
of matching rules.
llvm-svn: 353526
Make behavior of G_LOAD in widenScalar same as for G_ZEXTLOAD and
G_SEXTLOAD. That is perform widenScalarDst to size given by the target
and avoid additional checks in common code. Targets can reorder or add
additional rules in LegalizeRuleSet for the opcode to achieve desired
behavior.
Select extending load that does not have specified type of extension
into zero extending load.
Select truncating store that stores number of bytes indicated by size
in MachineMemoperand.
Differential Revision: https://reviews.llvm.org/D57454
llvm-svn: 353520
This is pretty much directly ported from SelectionDAG. Doesn't include
the shift by non-constant but known bits version, since there isn't a
globalisel version of computeKnownBits yet.
This shows a disadvantage of targets not specifically which type
should be used for the shift amount. If type 0 is legalized before
type 1, the operations on the shift amount type use the wider type
(which are also less likely to legalize). This can be avoided by
targets specifying legalization actions on type 1 earlier than for
type 0.
llvm-svn: 353455
Introduce a new function which handles instructions with multiple type
indices, but have the same number of vector elements.
Also legalize v2s16 shifts when applicable.
llvm-svn: 353432
The fewerElementsVectors implementation for load/stores
handles the scalar reduction case just as well, so drop
the redundant code in narrowScalar. This also introduces
support for narrowing irregular size breakdowns for
scalars.
llvm-svn: 353125
Don't handle vector conditions.
I think this can be merged in the future with
fewerElementsVectorSelect, although this becomes slightly tricky with
a vector condition.
llvm-svn: 353122
Try to use the underlying source registers.
This enables legalization in more cases where some irregular
operations are widened and others narrowed.
This seems to make the test_combines_2 AArch64 test worse, since the
MERGE_VALUES has multiple uses. Since this should be required for
legalization, a hasOneUse check is probably inappropriate (or maybe
should only be used if the merge is legal?).
llvm-svn: 353121
A number of of tests were using imm operands, not cimm. Since CSE
relies on the exact ConstantInt* pointer used, and implicit
conversions are generally evil, also enforce the bitsize of the types.
llvm-svn: 353113
For the scalar case only.
Also move the similar G_MERGE_VALUES handling to a separate function
and cleanup to make them look more similar.
llvm-svn: 352979
This was ignoring the memory size, and producing multiple loads/stores
if the operand size was different from the memory size.
I assume this is the intent of not having an explicit G_ANYEXTLOAD
(although I think that would probably be better).
llvm-svn: 352523
https://reviews.llvm.org/D57178
Now add a hook in TargetPassConfig to query if CSE needs to be
enabled. By default this hook returns false only for O0 opt level but
this can be overridden by the target.
As a consequence of the default of enabled for non O0, a few tests
needed to be updated to not use CSE (by passing in -O0) to the run
line.
reviewed by: arsenm
llvm-svn: 352126
It might be a bit nicer to use the fancy .legalIf and co. predicates,
but this was requiring more boilerplate and disables the coverage
assertions.
llvm-svn: 351886
There is a combine that was hiding these tests
not actually testing what they should be, although
they were producing the expected end result.
llvm-svn: 351698
This was crashing in the predicate function assuming the value
is a vector.
Copy more of what AArch64 uses. This probably needs more refinement
later, but I don't exactly understand what it means in some cases,
particularly since any legalization for these seems to be missing.
llvm-svn: 351693
I'm not entirely sure this is the correct thing
to do with the global isel philosophy, but I think
this is necessary to handle how differently SGPRs
are used normally vs. from a condition.
For example, it makes sense to allow a copy
from a VGPR to an SGPR, but it makes no sense
to allow a copy from VGPRs to SGPRs used as
select mask.
This avoids regbankselecting strange code with
a truncate feeding directly into a condition field.
Now a copy is forced from sgpr(s1) to vcc, which is
more sensible to handle.
Some of these issues could probably avoided with making enough
operations resulting in i1 illegal. I think we can't avoid
this register bank for legality.
For example, an i1 and where one source is from a truncate, and
one source is a compare needs some kind of copy inserted to
make sure both are in condition registers.
llvm-svn: 350611
If a copy was needed to handle the condition of brcond, it was being
inserted before the defining instruction. Add tests for iterator edge
cases.
I find the existing code here suspect for the case where it's looking
for terminators that modify the register. It's going to insert a copy
in the middle of the terminators, which isn't allowed (it might be
necessary to have a COPY_terminator if anybody actually needs this).
Also legalize brcond for AMDGPU.
llvm-svn: 350595
It seems better to avoid using the callback if possible since
there are coverage assertions which are disabled if this is used.
Also fix missing tests. Only test the legal cases since it seems
legalization for build_vector is quite lacking.
llvm-svn: 349878
This patch restricts the capability of G_MERGE_VALUES, and uses the new
G_BUILD_VECTOR and G_CONCAT_VECTORS opcodes instead in the appropriate places.
This patch also includes AArch64 support for selecting G_BUILD_VECTOR of <4 x s32>
and <2 x s64> vectors.
Differential Revisions: https://reviews.llvm.org/D53629
llvm-svn: 348788
Avoid using allocateKernArg / AssignFn. We do not want any
of the type splitting properties of normal calling convention
lowering.
For now at least this exists alongside the IR argument lowering
pass. This is necessary to handle struct padding correctly while
some arguments are still skipped by the IR argument lowering
pass.
llvm-svn: 336373
Summary:
We could split sizes that are not power of two into smaller sized
G_IMPLICIT_DEF instructions, but this ends up generating
G_MERGE_VALUES instructions which we then have to handle in the instruction
selector. Since G_IMPLICIT_DEF is really a no-op it's easier just to
keep everything that can fit into a register legal.
Reviewers: arsenm
Reviewed By: arsenm
Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, t-tye, llvm-commits
Differential Revision: https://reviews.llvm.org/D48777
llvm-svn: 336041
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
Note a normal select test is not currently possible because this
relies on input registers tracked in SIMachineFunctionInfo which
are not currently serializable in MIR, but this does work end-to-end
from the IR.
llvm-svn: 335490
Making sure we don't truncate / extend pointers, don't try to change
vector topology or bitcast vectors to scalars or back, and most
importantly, don't extend to a smaller type or truncate to a large
one.
Reviewers: qcolombet t.p.northover aditya_nandakumar
Reviewed By: qcolombet
Subscribers: rovka, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D46490
llvm-svn: 331718
As Roman Tereshin pointed out in https://reviews.llvm.org/D45541, the
-global-isel option is redundant when -run-pass is given. -global-isel sets up
the GlobalISel passes in the pass manager but -run-pass skips that entirely and
configures it's own pipeline.
llvm-svn: 331603
Matt Arsenault