We currently don't have a way to map to the equivalent intrinsic
opcode, so track immediate 0s in place of the address for the
selection to know to change the final opcode.
This reverts commit 9bca8fc4cf.
Rearrange handling to avoid changing the instruction in the case where
it's going to be erased and replaced with undef.
For normal loads, fully eliminate the load. For the TFE case, adjust
the dmask value in the instruction so the selector doesn't need to
handle it. For the TFE special case, I guess it would be possible to
replace the loaded data register with undef, but as-is this will start
treating it as a well defined value.
Trim elements that won't be written. The equivalent still needs to be
done for writes. Also start widening 3 elements to 4
elements. Selection will get the count from the dmask.
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.
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.
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.
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.
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.
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.
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.
Allows more flexible use of buildMerge in places where
use operands are available as SrcOp since it does not
require explicit conversion to Register.
Simplify code with new buildMerge.
Differential Revision: https://reviews.llvm.org/D74223
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.
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>.
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 is passed to legalizeCustom, but not intrinsic. Also remove the
MRI argument, since you can get that from the MachineIRBuilder.
I'm not sure why MachineIRBuilder has a private observer member, and
this is passed separately.
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.
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.
Bitcast only really applies between scalars and vectors. Implement as
an unmerge and remerge. The test needs to tolerate failure since one
of the unmerges currently fails to legalize.
The branch target needs to be changed depending on whether there is an
unconditional branch or not.
Loops also need to be similarly fixed, but compiling a simple testcase
end to end requires another set of patches that aren't upstream yet.
This should be legal, but will require future selection work. 16-bit
shift amounts were already removed from being legal, but this didn't
adjust the transformation rules.
This solves selection failures with generated selection patterns,
which would fail due to inferring the SGPR reg bank for virtual
registers with a set register class instead of VCC bank. Use
instruction selection would constrain the virtual register to a
specific class, so when the def was selected later the bank no longer
was set to VCC.
Remove the SCC reg bank. SCC isn't directly addressable, so it
requires copying from SCC to an allocatable 32-bit register during
selection, so these might as well be treated as 32-bit SGPR values.
Now any scalar boolean value that will produce an outupt in SCC should
be widened during RegBankSelect to s32. Any s1 value should be a
vector boolean during selection. This makes the vcc register bank
unambiguous with a normal SGPR during selection.
Summary of how this should now work:
- G_TRUNC is always a no-op, and never should use a vcc bank result.
- SALU boolean operations should be promoted to s32 in RegBankSelect
apply mapping
- An s1 value means vcc bank at selection. The exception is for
legalization artifacts that use s1, which are never VCC. All other
contexts should infer the VCC register classes for s1 typed
registers. The LLT for the register is now needed to infer the
correct register class. Extensions with vcc sources should be
legalized to a select of constants during RegBankSelect.
- Copy from non-vcc to vcc ensures high bits of the input value are
cleared during selection.
- SALU boolean inputs should ensure the inputs are 0/1. This includes
select, conditional branches, and carry-ins.
There are a few somewhat dirty details. One is that G_TRUNC/G_*EXT
selection ignores the usual register-bank from register class
functions, and can't handle truncates with VCC result banks. I think
this is OK, since the artifacts are specially treated anyway. This
does require some care to avoid producing cases with vcc. There will
also be no 100% reliable way to verify this rule is followed in
selection in case of register classes, and violations manifests
themselves as invalid copy instructions much later.
Standard phi handling also only considers the bank of the result
register, and doesn't insert copies to make the source banks
match. This doesn't work for vcc, so we have to manually correct phi
inputs in this case. We should add a verifier check to make sure there
are no phis with mixed vcc and non-vcc register bank inputs.
There's also some duplication with the LegalizerHelper, and some code
which should live in the helper. I don't see a good way to share
special knowledge about what types to use for intermediate operations
depending on the bank for example. Using the helper to replace
extensions with selects also seems somewhat awkward to me.
Another issue is there are some contexts calling
getRegBankFromRegClass that apparently don't have the LLT type for the
register, but I haven't yet run into a real issue from this.
This also introduces new unnecessary instructions in most cases, since
we don't yet try to optimize out the zext when the source is known to
come from a compare.
This would complain about invalid legalizer rules otherwise.
Mark some operations as unsupported for AMDGPU. This currently seems
to produce the same legalize error as when no rules are defined, but
eventually this should produce a proper user facing error.
There ended up being two result registers, which would fail on
select. It was really defing a new temp register in the correct def
position, instead of the correct result register.
Summary:
The only useful information the UndefValue conveys is the address space,
which MachinePointerInfo can represent directly without referring to an
IR value.
Reviewers: arsenm, rampitec
Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71838
Confusingly, the intrinsic operands do not match the
instruction/custom node. The order is shuffled, and the 3rd operand is
an immediate to select operands.
I'm not 100% sure I did this right, but fdiv still doesn't select end
to end and it will be easier to tell when it does. This at least
avoids an assertion in RegBankSelect and allows hitting the fallback
on selection.
Summary:
G_GEP is rather poorly named. It's a simple pointer+scalar addition and
doesn't support any of the complexities of getelementptr. I therefore
propose that we rename it. There's a G_PTR_MASK so let's follow that
convention and go with G_PTR_ADD
Reviewers: volkan, aditya_nandakumar, bogner, rovka, arsenm
Subscribers: sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, arphaman, Petar.Avramovic, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69734
Custom lower this to a target instruction with the merge operands. I
think it might be better to directly select this and emit a
REG_SEQUENCE, but this would be more work since it would require
splitting the tablegen patterns for these cases from the other
atomics.
There are 1024 bit register classes defined for AGPRs. Additionally
OpenCL defines vectors up to 16 x i64, and this helps those tests
legalize.
llvm-svn: 373350
This is sort of papering over the fact that we don't run a combiner
anywhere, but avoiding creating 2 instructions in the first place is
easy.
llvm-svn: 373293
This reverts r372314, reapplying r372285 and the commits which depend
on it (r372286-r372293, and r372296-r372297)
This was missing one switch to getTargetConstant in an untested case.
llvm-svn: 372338
This broke the Chromium build, causing it to fail with e.g.
fatal error: error in backend: Cannot select: t362: v4i32 = X86ISD::VSHLI t392, Constant:i8<15>
See llvm-commits thread of r372285 for details.
This also reverts r372286, r372287, r372288, r372289, r372290, r372291,
r372292, r372293, r372296, and r372297, which seemed to depend on the
main commit.
> Encode them directly as an imm argument to G_INTRINSIC*.
>
> Since now intrinsics can now define what parameters are required to be
> immediates, avoid using registers for them. Intrinsics could
> potentially want a constant that isn't a legal register type. Also,
> since G_CONSTANT is subject to CSE and legalization, transforms could
> potentially obscure the value (and create extra work for the
> selector). The register bank of a G_CONSTANT is also meaningful, so
> this could throw off future folding and legalization logic for AMDGPU.
>
> This will be much more convenient to work with than needing to call
> getConstantVRegVal and checking if it may have failed for every
> constant intrinsic parameter. AMDGPU has quite a lot of intrinsics wth
> immarg operands, many of which need inspection during lowering. Having
> to find the value in a register is going to add a lot of boilerplate
> and waste compile time.
>
> SelectionDAG has always provided TargetConstant for constants which
> should not be legalized or materialized in a register. The distinction
> between Constant and TargetConstant was somewhat fuzzy, and there was
> no automatic way to force usage of TargetConstant for certain
> intrinsic parameters. They were both ultimately ConstantSDNode, and it
> was inconsistently used. It was quite easy to mis-select an
> instruction requiring an immediate. For SelectionDAG, start emitting
> TargetConstant for these arguments, and using timm to match them.
>
> Most of the work here is to cleanup target handling of constants. Some
> targets process intrinsics through intermediate custom nodes, which
> need to preserve TargetConstant usage to match the intrinsic
> expectation. Pattern inputs now need to distinguish whether a constant
> is merely compatible with an operand or whether it is mandatory.
>
> The GlobalISelEmitter needs to treat timm as a special case of a leaf
> node, simlar to MachineBasicBlock operands. This should also enable
> handling of patterns for some G_* instructions with immediates, like
> G_FENCE or G_EXTRACT.
>
> This does include a workaround for a crash in GlobalISelEmitter when
> ARM tries to uses "imm" in an output with a "timm" pattern source.
llvm-svn: 372314
This needs special handling due to some subtargets that have a
nonstandard register layout for f16 vectors
Also reject some illegal types on other targets.
llvm-svn: 372293
Unlike SelectionDAG, treat this as a normally legalizable operation.
In SelectionDAG this is supposed to only ever formed if it's legal,
but I've found that to be restricting. For AMDGPU this is contextually
legal depending on whether denormal flushing is allowed in the use
function.
Technically we currently treat the denormal mode as a subtarget
feature, so custom lowering could be avoided. However I consider this
to be a defect, and this should be contextually dependent on the
controllable rounding mode of the parent function.
llvm-svn: 371800
f64 doesn't work yet because tablegen currently doesn't handlde
REG_SEQUENCE.
This does regress some multi use VALU fneg cases since now the
immediate remains in an SGPR, and more moves are used for legalizing
the xor. This is a SIFixSGPRCopies deficiency.
llvm-svn: 371540
There's still a lot more to do, but this handles decomposing due to
alignment. I've gotten it to the point where nothing crashes or
infinite loops the legalizer.
llvm-svn: 371533
Handle it the same way as G_BUILD_VECTOR_TRUNC. Arguably only
G_BUILD_VECTOR_TRUNC should be legal for this, but G_BUILD_VECTOR will
probably be more convenient in most cases.
llvm-svn: 371440
Treat this as legal on gfx9 since it can use S_PACK_* instructions for
this.
This isn't used by anything yet. The same will probably apply to
16-bit G_BUILD_VECTOR without the trunc.
llvm-svn: 371423
This is necessary for handling <3 x s16> on AMDGPU, assuming this
should be handled as 2 separate legalization actions. The alternative
would be for fewerElementsVector to handle 3->2.
llvm-svn: 369547
Summary:
Targets often have instructions that can sign-extend certain cases faster
than the equivalent shift-left/arithmetic-shift-right. Such cases can be
identified by matching a shift-left/shift-right pair but there are some
issues with this in the context of combines. For example, suppose you can
sign-extend 8-bit up to 32-bit with a target extend instruction.
%1:_(s32) = G_SHL %0:_(s32), i32 24 # (I've inlined the G_CONSTANT for brevity)
%2:_(s32) = G_ASHR %1:_(s32), i32 24
%3:_(s32) = G_ASHR %2:_(s32), i32 1
would reasonably combine to:
%1:_(s32) = G_SHL %0:_(s32), i32 24
%2:_(s32) = G_ASHR %1:_(s32), i32 25
which no longer matches the special case. If your shifts and extend are
equal cost, this would break even as a pair of shifts but if your shift is
more expensive than the extend then it's cheaper as:
%2:_(s32) = G_SEXT_INREG %0:_(s32), i32 8
%3:_(s32) = G_ASHR %2:_(s32), i32 1
It's possible to match the shift-pair in ISel and emit an extend and ashr.
However, this is far from the only way to break this shift pair and make
it hard to match the extends. Another example is that with the right
known-zeros, this:
%1:_(s32) = G_SHL %0:_(s32), i32 24
%2:_(s32) = G_ASHR %1:_(s32), i32 24
%3:_(s32) = G_MUL %2:_(s32), i32 2
can become:
%1:_(s32) = G_SHL %0:_(s32), i32 24
%2:_(s32) = G_ASHR %1:_(s32), i32 23
All upstream targets have been configured to lower it to the current
G_SHL,G_ASHR pair but will likely want to make it legal in some cases to
handle their faster cases.
To follow-up: Provide a way to legalize based on the constant. At the
moment, I'm thinking that the best way to achieve this is to provide the
MI in LegalityQuery but that opens the door to breaking core principles
of the legalizer (legality is not context sensitive). That said, it's
worth noting that looking at other instructions and acting on that
information doesn't violate this principle in itself. It's only a
violation if, at the end of legalization, a pass that checks legality
without being able to see the context would say an instruction might not be
legal. That's a fairly subtle distinction so to give a concrete example,
saying %2 in:
%1 = G_CONSTANT 16
%2 = G_SEXT_INREG %0, %1
is legal is in violation of that principle if the legality of %2 depends
on %1 being constant and/or being 16. However, legalizing to either:
%2 = G_SEXT_INREG %0, 16
or:
%1 = G_CONSTANT 16
%2:_(s32) = G_SHL %0, %1
%3:_(s32) = G_ASHR %2, %1
depending on whether %1 is constant and 16 does not violate that principle
since both outputs are genuinely legal.
Reviewers: bogner, aditya_nandakumar, volkan, aemerson, paquette, arsenm
Subscribers: sdardis, jvesely, wdng, nhaehnle, rovka, kristof.beyls, javed.absar, hiraditya, jrtc27, atanasyan, Petar.Avramovic, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61289
llvm-svn: 368487
AMDGPU sometimes has legal s16 and <2 x s16> operations, but all
registers are really 32-bit. An unmerge destination really should ben
widened to a 32-bit register. If widening a scalarizing vector with a
target size that matches the vector size, bitcast to integer and
extract the relevant bits with shifts.
I'm not sure if this is the right place for this. This could arguably
be part of widenScalar for the result. I also have a growing feeling
that we're missing a bitcast legalize action.
llvm-svn: 367604
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
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
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
Avoids using a plain unsigned for registers throughoug codegen.
Doesn't attempt to change every register use, just something a little
more than the set needed to build after changing the return type of
MachineOperand::getReg().
llvm-svn: 364191
This is incomplete, and ideally these would all be removed, but it's
better to localize them to the subtarget first with comments about
what they're for.
llvm-svn: 363902
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
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
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
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
Matt Arsenault