The DAG combiner logic to simplify AND masks in shift counts is invalid.
While it is true that the SystemZ shift instructions ignore all but the
low 6 bits of the shift count, it is still invalid to simplify the AND
masks while the DAG still uses the standard shift operators (which are
*not* defined to match the SystemZ instruction behavior).
Instead, this patch performs equivalent operations during instruction
selection. For completely removing the AND, this now happens via
additional DAG match patterns implemented by a multi-alternative
PatFrags. For simplifying a 32-bit AND to a 16-bit AND, the existing DAG
patterns were already mostly OK, they just needed an output XForm to
actually truncate the immediate value.
Unfortunately, the latter change also exposed a bug in TableGen: it
seems XForms are currently only handled correctly for direct operands of
the outermost operation node. This patch also fixes that bug by simply
recurring through the whole pattern. This should be NFC for all other
targets.
Differential Revision: https://reviews.llvm.org/D50096
llvm-svn: 338521
Currently, any attempt to define a PatFrag involving any floating-point
only (or vector only) node causes a hard assertion failure in TableGen
if the current target does not have any floating-point (or vector)
types.
This is annoying if you want to provide convenience fragments in common
code (e.g. include/llvm/Target/TargetSelectionDAG.td) that is parsed on
all platforms, including those that miss such types.
But really, there's no reason not accept this when parsing the fragment
-- of course it would be an error for such a target to actually *use*
such a fragment anywhere, but as long as it doesn't, I think TableGen
shouldn't error out.
The immediate cause of the assertion failure is the test inside the
ValidateOnExit destructor. This patch simply disables that check while
infering types during parsing of pattern fragments (only).
Reviewed By: hfinkel, kparzysz
Differential Revision: https://reviews.llvm.org/D48887
llvm-svn: 337023
A TableGen instruction record usually contains a DAG pattern that will
describe the SelectionDAG operation that can be implemented by this
instruction. However, there will be cases where several different DAG
patterns can all be implemented by the same instruction. The way to
represent this today is to write additional patterns in the Pattern
(or usually Pat) class that map those extra DAG patterns to the
instruction. This usually also works fine.
However, I've noticed cases where the current setup seems to require
quite a bit of extra (and duplicated) text in the target .td files.
For example, in the SystemZ back-end, there are quite a number of
instructions that can implement an "add-with-overflow" operation.
The same instructions also need to be used to implement just plain
addition (simply ignoring the extra overflow output). The current
solution requires creating extra Pat pattern for every instruction,
duplicating the information about which particular add operands
map best to which particular instruction.
This patch enhances TableGen to support a new PatFrags class, which
can be used to encapsulate multiple alternative patterns that may
all match to the same instruction. It operates the same way as the
existing PatFrag class, except that it accepts a list of DAG patterns
to match instead of just a single one. As an example, we can now define
a PatFrags to match either an "add-with-overflow" or a regular add
operation:
def z_sadd : PatFrags<(ops node:$src1, node:$src2),
[(z_saddo node:$src1, node:$src2),
(add node:$src1, node:$src2)]>;
and then use this in the add instruction pattern:
defm AR : BinaryRRAndK<"ar", 0x1A, 0xB9F8, z_sadd, GR32, GR32>;
These SystemZ target changes are implemented here as well.
Note that PatFrag is now defined as a subclass of PatFrags, which
means that some users of internals of PatFrag need to be updated.
(E.g. instead of using PatFrag.Fragment you now need to use
!head(PatFrag.Fragments).)
The implementation is based on the following main ideas:
- InlinePatternFragments may now replace each original pattern
with several result patterns, not just one.
- parseInstructionPattern delays calling InlinePatternFragments
and InferAllTypes. Instead, it extracts a single DAG match
pattern from the main instruction pattern.
- Processing of the DAG match pattern part of the main instruction
pattern now shares most code with processing match patterns from
the Pattern class.
- Direct use of main instruction patterns in InferFromPattern and
EmitResultInstructionAsOperand is removed; everything now operates
solely on DAG match patterns.
Reviewed by: hfinkel
Differential Revision: https://reviews.llvm.org/D48545
llvm-svn: 336999
So far, we've only handled special cases of PatFrag like ImmLeaf. This patch
adds support for the remaining cases using similar mechanisms.
Like most C++ code from SelectionDAG, GISel and DAGISel expect to operate on
different types and representations and as such the code is not compatible
between the two. It's therefore necessary to add an alternative implementation
in the GISelPredicateCode field.
The target test for this feature could easily be done with IntImmLeaf and this
would save on a little boilerplate. The reason I've chosen to implement this
using PatFrag.GISelPredicateCode and not IntImmLeaf is because I was unable to
find a rule that was blocked solely by lack of support for PatFrag predicates. I
found that the ones I investigated as being likely candidates for the test
were further blocked by other things.
llvm-svn: 334871
The return value of TreePatternNode::getChild is never null. This patch also
updates various places that use return values of getChild to also use
references. Those changes were suggested post-commit for D47463.
llvm-svn: 334764
By using std::shared_ptr for TreePatternNode, we can avoid leaking them.
Reviewers: craig.topper, dsanders, stoklund, tstellar, zturner
Reviewed By: dsanders
Differential Revision: https://reviews.llvm.org/D47463
llvm-svn: 333591
The DEBUG() macro is very generic so it might clash with other projects.
The renaming was done as follows:
- git grep -l 'DEBUG' | xargs sed -i 's/\bDEBUG\s\?(/LLVM_DEBUG(/g'
- git diff -U0 master | ../clang/tools/clang-format/clang-format-diff.py -i -p1 -style LLVM
- Manual change to APInt
- Manually chage DOCS as regex doesn't match it.
In the transition period the DEBUG() macro is still present and aliased
to the LLVM_DEBUG() one.
Differential Revision: https://reviews.llvm.org/D43624
llvm-svn: 332240
Summary:
r327219 added wrappers to std::sort which randomly shuffle the container before sorting.
This will help in uncovering non-determinism caused due to undefined sorting
order of objects having the same key.
To make use of that infrastructure we need to invoke llvm::sort instead of std::sort.
Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort.
Refer the comments section in D44363 for a list of all the required patches.
Reviewers: stoklund, kparzysz, dsanders
Reviewed By: dsanders
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45144
llvm-svn: 329451
This change deals with intrinsics with multiple outputs, for example load
instrinsic with address updated.
DAG selection for Instrinsics could be done either through source code or
tablegen. Handling all intrinsics in source code would introduce a huge chunk
of repetitive code if we have a large number of intrinsic that return multiple
values (see NVPTX as an example). While intrinsic class in tablegen supports
multiple outputs, tablegen only supports Intrinsics with zero or one output on
TreePattern. This appears to be a simple bug in tablegen that is fixed by this
change.
For Intrinsics defined as:
def int_xxx_load_addr_updated: Intrinsic<[llvm_i32_ty, llvm_ptr_ty], [llvm_ptr_ty, llvm_i32_ty], []>;
Instruction will be defined as:
def L32_X: Inst<(outs reg:$d1, reg:$d2), (ins reg:$s1, reg:$s2), "ld32_x $d1, $d2, $s2", [(set i32:$d1, i32:$d2, (int_xxx_load_addr_updated i32:$s1, i32:$s2))]>;
Patch by Wenbo Sun, thanks!
Differential Revision: https://reviews.llvm.org/D32888
llvm-svn: 321704
Allows preserving MachineMemOperands on intrinsics
through selection. For reasons I don't understand, this
is a static property of the pattern and the selector
deliberately goes out of its way to drop if not present.
Intrinsics already inherit from SDPatternOperator allowing
them to be used directly in instruction patterns. SDPatternOperator
has a list of SDNodeProperty, but you currently can't set them on
the intrinsic. Without SDNPMemOperand, when the node is selected
any memory operands are always dropped. Allowing setting this
on the intrinsics avoids needing to introduce another equivalent
target node just to have SDNPMemOperand set.
llvm-svn: 321212
Similar to r315841, GlobalISel and SelectionDAG require different code for the
common atomic predicates due to differences in the representation.
Even without that, differences in the IR (SDNode vs MachineInstr) require
differences in the C++ predicate.
This patch moves the implementation of the common atomic predicates related to
ordering into tablegen so that it can handle these differences.
It's NFC for SelectionDAG since it emits equivalent code and it's NFC for
GlobalISel since the rules involving the relevant predicates are still
rejected by the importer.
llvm-svn: 318102
Similar to r315841, GlobalISel and SelectionDAG require different code for the
common atomic predicates due to differences in the representation.
Even without that, differences in the IR (SDNode vs MachineInstr) require
differences in the C++ predicate.
This patch moves the implementation of the common atomic predicates related to
memory type into tablegen so that it can handle these differences.
It's NFC for SelectionDAG since it emits equivalent code and it's NFC for
GlobalISel since the rules involving the relevant predicates are still
rejected by the importer.
llvm-svn: 318095
Allow a pattern rewriter to be installed in CodeGenDAGPatterns and use it to
correct situations where SelectionDAG and GlobalISel disagree on
representation. For example, it would rewrite:
(sextload:i32 $ptr)<<unindexedload>><<sextload>><<sextloadi16>
to:
(sext:i32 (load:i16 $ptr)<<unindexedload>>)
I'd have preferred to replace the fragments and have the expansion happen
naturally as part of PatFrag expansion but the type inferencing system can't
cope with loads of types narrower than those mentioned in register classes.
This is because the SDTCisInt's on the sext constrain both the result and
operand to the 'legal' integer types (where legal is defined as 'a register
class can contain the type') which immediately rules the narrower types out.
Several targets (those with only one legal integer type) would then go on to
crash on the SDTCisOpSmallerThanOp<> when it removes all the possible types
for the result of the extend.
Also, improve isObviouslySafeToFold() slightly to automatically return true for
neighbouring instructions. There can't be any re-ordering problems if
re-ordering isn't happenning. We'll need to improve it further to handle
sign/zero-extending loads when the extend and load aren't immediate neighbours
though.
llvm-svn: 317971
Patch [1/5] in a series to add assembler/disassembler support for AArch64 SVE
unpredicated ADD/SUB instructions.
Patch by Sander De Smalen.
Reviewed by: rengolin
Differential Revision: https://reviews.llvm.org/D39087
llvm-svn: 317564
In type inference, an empty type set for a specific hw mode is not an
error. In earlier stages of the design it was, but having to use non-
parameterized types with target intrinsics necessarily led to type
contradictions: since the intrinsics used specific types, they were
only valid for a specific hw mode, and the resulting type set for other
modes ended up empty. To accommodate the existence of such intrinsics
individual type sets were allowed to be empty as long as not all sets
were empty.
llvm-svn: 315858
Summary:
GlobalISel and SelectionDAG require different code for the common
load/store predicates due to differences in the representation.
For example:
SelectionDAG: (load<signext,i8>:i32 GPR32:$addr) // The <> denote properties of the SDNode that are not printed in the DAG
GlobalISel: (G_SEXT:s32 (G_LOAD:s8 GPR32:$addr))
Even without that, differences in the IR (SDNode vs MachineInstr) require
differences in the C++ predicate.
This patch moves the implementation of the common load/store predicates
into tablegen so that it can handle these differences.
It's NFC for SelectionDAG since it emits equivalent code and it's NFC for
GlobalISel since the rules involving the relevant predicates are still
rejected by the importer.
Depends on D36618
Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar
Subscribers: llvm-commits, igorb
Differential Revision: https://reviews.llvm.org/D37443
Includes a partial revert of r315826 since this patch makes it necessary for
getPredCode() to return a std::string and getImmCode() should have the same
interface as getPredCode().
llvm-svn: 315841
Summary:
There's only a tablegen testcase for IntImmLeaf and not a CodeGen one
because the relevant rules are rejected for other reasons at the moment.
On AArch64, it's because there's an SDNodeXForm attached to the operand.
On X86, it's because the rule either emits multiple instructions or has
another predicate using PatFrag which cannot easily be supported at the
same time.
Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar
Reviewed By: qcolombet
Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D36569
llvm-svn: 315761
Summary:
The purpose of this patch is to expose more information about ImmLeaf-like
PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf
could only be used to test int64_t's produced by sign-extending an APInt.
Other tests on immediates had to use the generic PatLeaf and extract the
constant using C++.
With this patch, tablegen will know how to generate predicates for APInt,
and APFloat. This will allow it to 'do the right thing' for both SelectionDAG
and GlobalISel which require different methods of extracting the immediate
from the IR.
This is NFC for SelectionDAG since the new code is equivalent to the
previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1
for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new
subclasses will require a significant re-factor of FastISel.
For GlobalISel, it's currently NFC because the relevant code to import the
affected rules is not yet present. This will be added in a later patch.
Depends on D36086
Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar
Reviewed By: qcolombet
Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D36534
llvm-svn: 315747
I'm about to commit a patch that makes them necessary for getPredCode() and
it would be strange for getPredCode() and getImmCode() to require different
usage.
llvm-svn: 315733
Avoid unnecessary std::string creations in the TreePredicateFn getters and in CodeGenDAGPatterns::getSDNodeNamed
Differential Revision: https://reviews.llvm.org/D38624
llvm-svn: 315148
Also add operator<< for use with raw_ostream to InfoByHwMode and its
derived classes.
Recommitting r313989 with the fix for unresolved references: explicitly
define the operator<< in namespace llvm.
llvm-svn: 314004
This changes some STL data types to corresponding LLVM
data types that have better performance characteristics.
Differential Revision: https://reviews.llvm.org/D37957
llvm-svn: 313783
Add some member types to MachineValueTypeSet::const_iterator so that
iterator_traits can work with it.
Improve TableGen performance of -gen-dag-isel (motivated by X86 backend)
The introduction of parameterized register classes in r313271 caused the
matcher generation code in TableGen to run much slower, particularly so
in the unoptimized (debug) build. This patch recovers some of the lost
performance.
Summary of changes:
- Cache the set of legal types in TypeInfer::getLegalTypes. The contents
of this set do not change.
- Add LLVM_ATTRIBUTE_ALWAYS_INLINE to several small functions. Normally
this would not be necessary, but in the debug build TableGen is not
optimized, so this helps a little bit.
- Add an early exit from TypeSetByHwMode::operator== for the case when
one or both arguments are "simple", i.e. only have one mode. This
saves some time in GenerateVariants.
- Finally, replace the underlying storage type in TypeSetByHwMode::SetType
with MachineValueTypeSet based on std::array instead of std::set.
This significantly reduces the number of memory allocation calls.
I've done a number of experiments with the underlying type of InfoByHwMode.
The type is a map, and for targets that do not use the parameterization,
this map has only one entry. The best (unoptimized) performance, somewhat
surprisingly came from std::map, followed closely by std::unordered_map.
DenseMap was the slowest by a large margin.
Various hand-crafted solutions (emulating enough of the map interface
not to make sweeping changes to the users) did not yield any observable
improvements.
llvm-svn: 313660
The introduction of parameterized register classes in r313271 caused the
matcher generation code in TableGen to run much slower, particularly so
in the unoptimized (debug) build. This patch recovers some of the lost
performance.
Summary of changes:
- Cache the set of legal types in TypeInfer::getLegalTypes. The contents
of this set do not change.
- Add LLVM_ATTRIBUTE_ALWAYS_INLINE to several small functions. Normally
this would not be necessary, but in the debug build TableGen is not
optimized, so this helps a little bit.
- Add an early exit from TypeSetByHwMode::operator== for the case when
one or both arguments are "simple", i.e. only have one mode. This
saves some time in GenerateVariants.
- Finally, replace the underlying storage type in TypeSetByHwMode::SetType
with MachineValueTypeSet based on std::array instead of std::set.
This significantly reduces the number of memory allocation calls.
I've done a number of experiments with the underlying type of InfoByHwMode.
The type is a map, and for targets that do not use the parameterization,
this map has only one entry. The best (unoptimized) performance, somewhat
surprisingly came from std::map, followed closely by std::unordered_map.
DenseMap was the slowest by a large margin.
Various hand-crafted solutions (emulating enough of the map interface
not to make sweeping changes to the users) did not yield any observable
improvements.
llvm-svn: 313647
Ulrich Weigand