This patch relands https://reviews.llvm.org/D104454, but fixes some failing
builds on Mac OS which apparently has a different definition for size_t,
that caused 'ambiguous operator overload' for the implicit conversion
of TypeSize to a scalar value.
This reverts commit b732e6c9a8.
To reflect that the size may be scalable, a TypeSize is returned
instead of an unsigned. In places where the result is used,
it currently relies on an implicit cast of TypeSize -> uint64_t,
which asserts that the type is not scalable.
This patch is NFC for fixed-width vectors.
Reviewed By: aemerson
Differential Revision: https://reviews.llvm.org/D104454
This also adds new interfaces for the fixed- and scalable case:
* LLT::fixed_vector
* LLT::scalable_vector
The strategy for migrating to the new interfaces was as follows:
* If the new LLT is a (modified) clone of another LLT, taking the
same number of elements, then use LLT::vector(OtherTy.getElementCount())
or if the number of elements is halfed/doubled, it uses .divideCoefficientBy(2)
or operator*. That is because there is no reason to specifically restrict
the types to 'fixed_vector'.
* If the algorithm works on the number of elements (as unsigned), then
just use fixed_vector. This will need to be fixed up in the future when
modifying the algorithm to also work for scalable vectors, and will need
then need additional tests to confirm the behaviour works the same for
scalable vectors.
* If the test used the '/*Scalable=*/true` flag of LLT::vector, then
this is replaced by LLT::scalable_vector.
Reviewed By: aemerson
Differential Revision: https://reviews.llvm.org/D104451
This patch aims to add the scalable property to LLT. The rest of the
patch-series changes the interfaces to take/return ElementCount and
TypeSize, which both have the ability to represent the scalable property.
The changes are mostly mechanical and aim to be non-functional changes
for fixed-width vectors.
For scalable vectors some unit tests have been added, but no effort has
been put into making any of the GlobalISel algorithms work with scalable
vectors yet. That will be left as future work.
The work is split into a series of 5 patches to make reviews easier.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D104450
This uses to be how predicates were handled prior to HwMode being
added. When the Predicates were converted to a std::vector it
significantly increased the cost of a compare in GenerateVariants.
Since ListInit's are uniquified by tablegen, we can use a simple
pointer comparison to check for identical lists.
In order to store the HwMode, we now add a separate string to
PatternToMatch. This will be appended separately to the predicate
string in getPredicateCheck. A new getPredicateRecords is added
to allow GlobalISel and getPredicateCheck to both get the sorted
list of Records. GlobalISel was ignoring any HwMode predicates
before and still is.
There is one slight change here, ListInits with different predicate
orders aren't sorted so the filtering in GenerateVariants might
fail to detect two isomorphic patterns with different predicate
orders. This doesn't seem to be happening in tree today.
My hope is this will allow us to remove all the BitVector tracking
in GenerateVariants that was making up for predicates beeing
expensive to compare. There's a decent amount of heap allocations
there on large targets like X86, AMDGPU, and RISCV.
Differential Revision: https://reviews.llvm.org/D100691
When GlobalISelEmitter::emitCxxPredicateFns emitted code for MI
predicates it used "PatFrag" when searching for definitions. With
this patch it will search for all "PatFrags" instead. Since PatFrag
derives from PatFrags the difference is that we now include all
definitions using PatFrags directly as well. Thus making it possible
to use GISelPredicateCode together with a PatFrags definition.
It might be noted that the matcher code was emitted also for PatFrags
in the past. But then one ended up with errors since the custom code
in testMIPredicate_MI was missing.
Differential Revision: https://reviews.llvm.org/D98486
To do this while supporting the existing functionality in SelectionDAG of using
PGO info, we add the ProfileSummaryInfo and LazyBlockFrequencyInfo analysis
dependencies to the instruction selector pass.
Then, use the predicate to generate constant pool loads for f32 materialization,
if we're targeting optsize/minsize.
Differential Revision: https://reviews.llvm.org/D97732
Allow different GICustomOperandRenderers to use the same RendererFn.
This avoids the need for targets to define a bunch of identical C++
renderer functions with different names.
Without this fix TableGen would have emitted code that tried to define
the GICR enumeration with duplicate enumerators.
Differential Revision: https://reviews.llvm.org/D96587
Before this patch there was generic mapping from vector_extract
to G_EXTRACT_VECTOR_ELT added in SelectionDAGCompat.td. That
mapping is now replaced by a mapping from extractelt instead.
The reasoning is that vector_extract is marked as deprecated,
so it is assumed that a majority of targets will use extractelt
and not vector_extract (and that the long term solution for all
targets would be to use extractelt).
Targets like AArch64 that still use vector_extract can add an
additional mapping from the deprecated vector_extract as target
specific tablegen definitions. Such a mapping is added for AArch64
in this patch to avoid breaking tests.
When adding the extractelt => G_EXTRACT_VECTOR_ELT mapping we
triggered some new code paths in GlobalISelEmitter, ending up in
an assert when trying to import a pattern containing EXTRACT_SUBREG
for ARM. Therefore this patch also adds a "failedImport" warning
for that situation (instead of hitting the assert).
Differential Revision: https://reviews.llvm.org/D93416
TableGen would pick the largest RC for constraining the operands, which
could potentially be an unallocatable RC. This patch removes selection
of unallocatable RCs.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D93945
The companion RFC (http://lists.llvm.org/pipermail/llvm-dev/2020-October/145850.html) gives lots of details on the overall strategy, but we summarize it here:
LLVM IR involving vector types is going to be selected using pseudo instructions (only MachineInstr). These pseudo instructions contain dummy operands to represent the vector type being operated and the vector length for the operation.
These two dummy operands, as set by instruction selection, will be used by the custom inserter to prepend every operation with an appropriate vsetvli instruction that ensures the vector architecture is properly configured for the operation. Not in this patch: later passes will remove the redundant vsetvli instructions.
Register classes of tuples of vector registers are used to represent vector register groups (LMUL > 1).
Those pseudos are eventually lowered into the actual instructions when emitting the MCInsts.
About the patch:
Because there is a bit of initial infrastructure required, this is the minimal patch that allows us to select instructions for 3 LLVM IR instructions: load, add and store vectors of integers. LLVM IR operations have "whole-vector" semantics (as in they generate values for all the elements).
Later patches will extend the information represented in TableGen.
Authored-by: Roger Ferrer Ibanez <rofirrim@gmail.com>
Co-Authored-by: Evandro Menezes <evandro.menezes@sifive.com>
Co-Authored-by: Craig Topper <craig.topper@sifive.com>
Differential Revision: https://reviews.llvm.org/D89449
Tablegen seg faulted when parsing a Pat where the destination part has
no output (zero instruction), due to a register class lookup using
nullptr.
Reviewed By: Paul-C-Anagnostopoulos
Differential Revision: https://reviews.llvm.org/D90829
Some of these were found by running clang-format over the generated
code, although that complains about far more issues than I have fixed
here.
Differential Revision: https://reviews.llvm.org/D90937
When nesting INSERT_SUBREG and EXTRACT_SUBREG, GlobalISelEmitter would
fail to find the register class of the nested node. This patch fixes
that for registers with subregs.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D88487
When generating matching tables for GlobalISel, TableGen would output
"::zero_reg" whenever encountering the zero_reg, which in turn would
result in compilation error. This patch fixes that by instead outputting
NoRegister (== 0), which is the same result that TableGen produces when
generating matching tables for ISelDAG.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D86215
The "name" of a non-leaf complex pattern (MY_PAT $op1, $op2) is
"MY_PAT:op1:op2" and the ones with same "name" represent same operand.
Add 'same operand check' for this case.
Differential Revision: https://reviews.llvm.org/D87351
Predicates with 'let PredicateCodeUsesOperands = 1' want to examine
matched operands. When we encounter predicate code that uses operands,
analyze its named operand arguments and create a map between argument
index and name. Later, when leaf node with name is encountered, emit
GIM_RecordNamedOperand that will store that operand at its argument
index in operand list. This operand list will be an argument to c++
code of the predicate.
Differential Revision: https://reviews.llvm.org/D87285
When optimizing the table, PointerToAnyOperandMatchers would be
incorrectly reported as identical even though they have different
SizeInBits values. This bug was due to failing to overload the
isIdentical() method, which this patch addresses.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D86199
This is to initially handleg immAllOnesV, which should match
G_BUILD_VECTOR or G_BUILD_VECTOR_TRUNC. In the future, it could be
used for other patterns cases that map to multiple G_* instructions,
such as G_ADD and G_PTR_ADD.
This patch fixes a bug which skipped
adding predicate matcher for a pattern in many cases.
For example, if predicate is Load and
its memoryVT is non-null then the loop
continues and never reaches to the end which
adds the predicate matcher. This patch moves the
matcher addition to the top of the loop
so that it gets added regardless of contextual checks
later in the loop.
Other way to fix this issue is to remove all "continue" statements
in checks and let the loop continue till end.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D83034
These should really match either G_BUILD_VECTOR or
G_BUILD_VECTOR_TRUNC, but there doesn't seem to be an existing
mechanism for matching alternative opcodes. There is GIM_SwitchOpcode,
but it seems to assume it's oly only used for matcher optimization.
I could also omit any opcode check and rely on the matcher directly
checking the opcode, but the table optimizer currently assumes there
has to be an opcode check.
Also doesn't try to handle undef elements like the DAG version.
ISD::ATOMIC_STORE arbitrarily has the operands in the opposite order
from regular ISD::STORE, which always introduced an annoying
duplication of patterns to handle both cases. Since in GlobalISel
there's just the one G_STORE, we need to swap the operands to
correctly emit the type check for the pointer operand.
Some work started in 20aafa3156 to
migrate SelectionDAG to use ISD::STORE for atomics, but that work
seems to have stalled. Since this is the pretty much the last
operation which matters which isn't supported for AMDGPU, use this
compatibility hack to unblock declaring it functionally complete.
Not sure what's going on with the pending_phis AArch64 test. It seems
it didn't always use atomics, and I'm not sure what it was originally
testing matters anymore.
The DAG behavior allows matchching input patterns with a single result
to the first result of an output instruction that defines multiple
results. The remaining defs are implicitly dead.
This starts to fix using manual selection for AMDGPU add/sub (although
it's still needed, mostly because it's also still needed for
G_PTR_ADD).
Currently custom code predicates can only really be used for
contextless checks tied to a single instruction (e.g. check the def
for hasOneUse). If you do want to inspect the input instructions in
the source pattern, you cannot without re-verifying the opcode and
type checks implied by the patterns, since this check was emitted
before any operand constraints. Really, these are pattern level
predicates that implicitly depend on the instruction and operand
checks.
Introduce a filtering function so the custom predicate is emitted
last. I'm not sure this is the most elegant solution. It seems like
this is really a different thing from the InstructionMatcher/IPM_
predicate kinds. I initially tried keeping this in a separate
predicate list, but that also seemed awkward.
This only half fixes the problem I'm trying to solve. The AMDGPU
pattern I'm attempting to port also uses the PredicateCodeUsesOperands
feature to allow checks on the source operands when the input pattern
is commuted. Really the emitter should reject the pattern since it
doesn't handle this case, but at this point it would be more
productive to just implement this.
This was emitting the raw value for the reg class ID with a comment
for the actual class name. Switch to emitting the qualified enum name
instead, which obviates the need for the comment and also helps keep
the lit tests on the emitter output more stable.
This was hitting the default instruction constraint code which uses
the register classes in the instruction def, which REG_SEQUENCE does
not have.
Fixes not constraining the register class for AMDGPU fneg/fabs
patterns, which would fail when the use was another generic,
unconstrained instruction.
Another oddity I noticed is that the temporary registers are created
with an unnecessary, but incorrect 16-bit LLT but this shouldn't
matter.
I'm also still unclear why root and sub-instructions have to be
handled differently.
This is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
This previously only handled EXTRACT_SUBREGs from leafs, such as
operands directly in the original output. Handle extracting from a
result instruction.
These return temporary Optional<> values which are immediately
destroyed. I'm not sure why no sanitizers seem to have caught this,
but I encountered crashes on these in a future patch.
The maps for dealing with the relationships between different register
classes and subregister indexes rely on unique pointers for every
class/index. By constructing a second copy of CodeGenRegBank, two
different pointer values existed for a given subregister depending on
where you were querying.
Use the existing CodeGenRegBank owned by the CodeGenTarget instead of
constructing a second copy. This avoids incorrectly failing map
lookups in a future change.
For arguments that are not expected to be materialized with
G_CONSTANT, this was emitting predicates which could never match. It
was first adding a meaningless LLT check, which would always fail due
to the operand not being a register.
Infer the cases where a literal should check for an immediate operand,
instead of a register This avoids needing to invent a special way of
representing timm literal values.
Also handle immediate arguments in GIM_CheckLiteralInt. The comments
stated it handled isImm() and isCImm(), but that wasn't really true.
This unblocks work on the selection of all of the complicated AMDGPU
intrinsics in future commits.
The current implementation assumes there is an instruction associated
with the transform, but this is not the case for
timm/TargetConstant/immarg values. These transforms should directly
operate on a specific MachineOperand in the source
instruction. TableGen would assert if you attempted to define an
equivalent GISDNodeXFormEquiv using timm when it failed to find the
instruction matcher.
Specially recognize SDNodeXForms on timm, and pass the operand index
to the render function.
Ideally this would be a separate render function type that looks like
void renderFoo(MachineInstrBuilder, const MachineOperand&), but this
proved to be somewhat mechanically painful. Add an optional operand
index which will only be passed if the transform should only look at
the one source operand.
Theoretically it would also be possible to only ever pass the
MachineOperand, and the existing renderers would check the parent. I
think that would be somewhat ugly for the standard usage which may
want to inspect other operands, and I also think MachineOperand should
eventually not carry a pointer to the parent instruction.
Use it in one sample pattern. This isn't a great example, since the
transform exists to satisfy DAG type constraints. This could also be
avoided by just changing the MachineInstr's arbitrary choice of
operand type from i16 to i32. Other patterns have nontrivial uses, but
this serves as the simplest example.
One flaw this still has is if you try to use an SDNodeXForm defined
for imm, but the source pattern uses timm, you still see the "Failed
to lookup instruction" assert. However, there is now a way to avoid
it.
Sander de Smalen