Summary:
This patch adds a new intrinsic and builtin function mirroring the
existing `__builtin_readcyclecounter`. The difference is that this
implementation targets a separate counter that some targets have which
returns a fixed frequency clock that can be used to determine elapsed
time, this is different compared to the cycle counter which often has
variable frequency.
This patch only adds support for the NVPTX and AMDGPU targets.
This is done as a new and separate builtin rather than an argument to
`readcyclecounter` to avoid needing to change existing code and to make
the separation more explicit.
This tries to fix a bug by resolving a few FIXMEs. The bug is that
`EraseInstAction` is emitted after emitting the _first_ `BuildMIAction`,
which is too early because the erased instruction may still be used by
subsequent `BuildMIAction`s (in particular, by `CopyRenderer`).
An example of the bug (from `match-table-operand-types.td`):
```
def InstTest0 : GICombineRule<
(defs root:$a),
(match (G_MUL i32:$x, i32:$b, i32:$c),
(G_MUL $a, i32:$b, i32:$x)),
(apply (G_ADD i64:$tmp, $b, i32:$c),
(G_ADD i8:$a, $b, i64:$tmp))>;
GIR_EraseFromParent, /*InsnID*/0,
GIR_BuildMI, /*InsnID*/1, /*Opcode*/GIMT_Encode2(TargetOpcode::G_ADD),
GIR_Copy, /*NewInsnID*/1, /*OldInsnID*/0, /*OpIdx*/0, // a
GIR_Copy, /*NewInsnID*/1, /*OldInsnID*/0, /*OpIdx*/1, // b
GIR_AddSimpleTempRegister, /*InsnID*/1, /*TempRegID*/0,
```
Here, the root instruction is destroyed before copying its operands ('a'
and 'b') to the new instruction.
The solution is to emit `EraseInstAction` for the root instruction as
the last action in the emission pipeline.
The intrinsics get_fpenv, set_fpenv and reset_fpenv in this change are
implemented as calls to math library functions. Target specific lowering
will be implemented later on.
- Instead of checking the default ops directly, this change queries DAG
default operands collected during patterns reading. It does not only
simplify the code but also handle few cases where integer values are
converted from convertible types, such as 'bits'.
- A test case is added GlobalISelEmitter.td as the regression test of
default 'bits' values.
Most users of AddImm and CheckConstantInt only use 1 byte immediates, so
I added an opcode variants for those. That way all those instructions
save 7 bytes.
Also added an opcode for AddTempRegister for the cases where there are
no register flags.
Space savings:
- AMDGPUGenGlobalISel: 470180 bytes to 422564 (-10%)
- AArch64GenGlobalISel.inc: 383893 bytes to 374046
When importing instruction selection patterns into GlobalISel, the
operands matched in the "source" DAG are copied into corresponding
operands of the "destination" DAG according to their names (such as Rd).
If multiple operands in the source DAG share the same name, a
GIM_CheckIsSameOperand predicate makes instruction selector check the
corresponding operands for equality (at compiler run-time) as part of
matching the source pattern.
The Def operands of the root node of the destination DAG are handled
specially. The operands of the instruction corresponding to the root
node are taken and GIM_CheckRegBankForClass predicates are
tablegen-erated accordingly. If by coincidence the Def operand in
question has the same name as one of the named operands in the pattern,
a GIM_CheckIsSameOperand predicate is automatically added that is likely
to prevent matching the source of otherwise applicable selection pattern
at compiler run-time.
This patch mangles the Def operand names taken from the instruction
corresponding to the root of the destination DAG (for example, "Rd"
becomes "DstI[Rd]") preventing unexpected name clashes with pattern's
named operands.
The patch consists of three sets of changes:
* changes to the GlobalISelEmitter.cpp file are the actual fix
* a test case is added to GlobalISelEmitter.td file as a regression test
* everything else is the biggest and least interesting part - updates to
the existing test cases: renames of the form Rd -> DstI[Rd] inside the
inline comments in tablegen-erated code
The MatchTableExecutor did not use the MachineIRBuilder but instead
created instructions ad-hoc.
Making it use a Builder has the benefit that any observer added by a
combine is now notified when instructions are created by MIR patterns.
Another benefit is that it allows me to improve how constants are
created in apply MIR patterns.
`MachineIRBuilder::buildConstant` automatically handles splats for us,
this means that we may change `addCImm` to use that and handle vector
cases automatically.
PredicateBitset currently uses std::bitset, but std::bitset doesn't
have a constexpr constructor or any constexpr methods until C++23.
Each target that supports GlobalIsel has as an array of PredicateBitset
objects that currently use a global constructor.
SubtargetFeature used by the MC layer for feature bits, has its own
implementation of std::bitset that has constexpr constructor and methods
that provides all the capabilities that PredicateBitset needs.
This patch copies the implementation from SubtargetFeature, makes
it a template class, and puts it in ADT. I'll migrate SubtargetFeature
in a separate patch. Adapting all existing users to it being a template
was distracting from the goal of this patch.
This reduces the binary size of llc built with gcc 8.5.0 on my local
build by ~15k.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D158576
Adds a new backend to power the GISel Combiners using the InstructionSelector's match tables.
This does not depend on any of the data structures created for the current combiner and is intended to replace it entirely.
See the RFC for more details: https://discourse.llvm.org/t/rfc-matchtable-based-globalisel-combiners/71457/6
Note: this would replace D141135.
Reviewed By: aemerson, arsenm
Differential Revision: https://reviews.llvm.org/D153757
Move all of the reusable logic out of `GlobalISelEmitter.cpp` into a `GlobalISelMatchTableExecutorEmitter` class so the future combiner backend can use it as well.
Depends on D153755
Reviewed By: aemerson
Differential Revision: https://reviews.llvm.org/D153756
Makes `InstructionSelector.h`/`InstructionSelectorImpl.h` generic so the match tables can also be used for the combiner.
Some notes:
- Coverage was made an optional parameter of `executeMatchTable`, combines won't use it for now.
- `GIPFP_` -> `GICXXPred_` so it's more generic. Those are just C++ predicates and aren't PatFrag-specific.
- Pass the MatcherState directly to testMIPredicate_MI, the combiner will need it.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D153755
Without it, the count of renderer functions is inaccurate and, in some
edge cases (like the patterns added in D134354), we can actually
go out of bounds (run out of pre-allocated renderer function spaces
in the GISel state)
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D134861
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
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
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
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.
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.
This requires std::intializer_list to be a literal type, which it is
starting with C++14. The downside is that std::bitset is still not
constexpr-friendly so this change contains a re-implementation of most
of it.
Shrinks clang by ~60k.
llvm-svn: 369847
Factor out commonly-used target code from the GlobalISelEmitter tests into
a GlobalISelEmitterCommon.td file. This is tested by the original
GlobalISelEmitter.td test.
This reduces the amount of boilerplate code necessary for tests like this.
Differential Revision: https://reviews.llvm.org/D65777
llvm-svn: 368757
Summary:
This adds support for defining patterns for global isel using pointer
types, for example:
def : Pat<(load GPR32:$src),
(p1 (LOAD GPR32:$src))>;
DAGISelEmitter will ignore the pointer information and treat these
types as integers with the same bit-width as the pointer type.
Reviewers: dsanders, rtereshin, arsenm
Reviewed By: arsenm
Subscribers: Petar.Avramovic, wdng, rovka, kristof.beyls, jfb, volkan, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D57065
llvm-svn: 354510
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
This patch continues a series of patches started by r332907 (reapplied
as r332917).
In this commit we move register bank checks back from epilogue of
every rule matcher to a position locally close to the rest of the
checks for a particular (nested) instruction.
This increases the number of common conditions within 2nd level
groups.
This is expected to decrease time GlobalISel spends in its
InstructionSelect pass by about 2% for an -O0 build as measured on
sqlite3-amalgamation (http://sqlite.org/download.html) targeting
AArch64 (cross-compile on x86).
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 333144
This patch continues a series of patches started by r332907 (reapplied
as r332917).
In this commit we greedily stuff 2nd level GroupMatcher's common
conditions with as many predicates as possible. This is purely
post-processing and it doesn't change which rules are put into the
groups in the first place: that decision is made by looking at the
first common predicate only.
The compile time improvements are minor and well within error margin,
however, it's highly improbable that this transformation could
pessimize performance, thus I'm still committing it for potential
gains for targets not implementing GlobalISel yet and out of tree
targets.
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 333139
This patch continues a series of patches started by r332907 (reapplied
as r332917)
In this commit we sort type checks towards the beginning of every rule
within the MatchTable as they fail often and it's best to fail early.
This is expected to decrease time GlobalISel spends in its
InstructionSelect pass by roughly 7% for an -O0 build as measured on
sqlite3-amalgamation (http://sqlite.org/download.html) targeting
AArch64. The amalgamation is a large single-file C-source that makes
compiler backend performance improvements to stand out from frontend.
It's also a part of CTMark.
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 333114
This patch continues a series of patches started by r332907 (reapplied
as r332917)
In this commit we start grouping rules with common first condition on
the second level of the table.
This is expected to decrease time GlobalISel spends in its
InstructionSelect pass by roughly 13% for an -O0 build as measured on
sqlite3-amalgamation (http://sqlite.org/download.html) targeting
AArch64.
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 333053
This patch continues a series of patches started by r332907 (reapplied
as r332917)
In this commit we introduce a new matching opcode GIM_SwitchOpcode
that implements a jump table over opcodes and start emitting them for
root instructions.
This is expected to decrease time GlobalISel spends in its
InstructionSelect pass by roughly 20% for an -O0 build as measured on
sqlite3-amalgamation (http://sqlite.org/download.html) targeting
AArch64.
To some degree, we assume here that the opcodes form a dense set,
which is true at the moment for all upstream targets given the
limitations of our rule importing mechanism.
It might not be true for out of tree targets, specifically due to
pseudo's. If so, we might noticeably increase the size of the
MatchTable with this patch due to padding zeros. This will be
addressed later.
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 333017
This patch continues a series of patches that decrease time spent by
GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds
for large inputs as measured on sqlite3-amalgamation
(http://sqlite.org/download.html) targeting AArch64.
This commit specifically removes number of operands checks that are
redundant if the instruction's opcode already guarantees that number
of operands (or more), and also avoids any kind of checks on a def
operand of a nested instruction as everything about it was already
checked at its use.
The expected performance implication is about 3% off InstructionSelect
comparing to the baseline (before the series of patches)
This patch also contains a bit of NFC changes required for further
patches in the series.
Every commit planned shares the same Phabricator Review.
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 332945
Apparently the compile time problem was caused by the fact that not
all compilers / STL implementations can automatically convert
std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully)
by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to
std::unique_ptr<Base> conversion instead.
llvm-svn: 332917
This patch starts a series of patches that decrease time spent by
GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds
for large inputs as measured on sqlite3-amalgamation
(http://sqlite.org/download.html) targeting AArch64.
The performance improvements are achieved solely by reducing the
number of matching GIM_* opcodes executed by the MatchTable's
interpreter during the selection by approx. a factor of 30, which also
brings contribution of this particular part of the selection process
to the overall runtime of InstructionSelect pass down from approx.
60-70% to 5-7%, thus making further improvements in this particular
direction not very profitable.
The improvements described above are expected for any target that
doesn't have many complex patterns. The targets that do should
strictly benefit from the changes, but by how much exactly is hard to
estimate beforehand. It's also likely that such target WILL benefit
from further improvements to MatchTable, most likely the ones that
bring it closer to a perfect decision tree.
This commit specifically is rather large mostly NFC commit that does
necessary preparation work and refactoring, there will be a following
series of small patches introducing a specific optimization each
shortly after.
This commit specifically is expected to cause a small compile time
regression (around 2.5% of InstructionSelect pass time), which should
be fixed by the next commit of the series.
Every commit planned shares the same Phabricator Review.
Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar
Reviewed By: qcolombet
Subscribers: rovka, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D44700
llvm-svn: 332907
Summary:
Previously, a extending load was represented at (G_*EXT (G_LOAD x)).
This had a few drawbacks:
* G_LOAD had to be legal for all sizes you could extend from, even if
registers didn't naturally hold those sizes.
* All sizes you could extend from had to be allocatable just in case the
extend went missing (e.g. by optimization).
* At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we
improve optimization of extends and truncates, this legality requirement
would spread without considerable care w.r.t when certain combines were
permitted.
* The SelectionDAG importer required some ugly and fragile pattern
rewriting to translate patterns into this style.
This patch changes the representation to:
* (G_[SZ]EXTLOAD x)
* (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits()
which resolves these issues by allowing targets to work entirely in their
native register sizes, and by having a more direct translation from
SelectionDAG patterns.
Each extending load can be lowered by the legalizer into separate extends
and loads, however a target that supports s1 will need the any-extending
load to extend to at least s8 since LLVM does not represent memory accesses
smaller than 8 bit. The legalizer can widenScalar G_LOAD into an
any-extending load but sign/zero-extending loads need help from something
else like a combiner pass. A follow-up patch that adds combiner helpers for
for this will follow.
The new representation requires that the MMO correctly reflect the memory
access so this has been corrected in a couple tests. I've also moved the
extending loads to their own tests since they are (mostly) separate opcodes
now. Additionally, the re-write appears to have invalidated two tests from
select-with-no-legality-check.mir since the matcher table no longer contains
loads that result in s1's and they aren't legal in AArch64 anymore.
Depends on D45540
Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar
Reviewed By: rtereshin
Subscribers: javed.absar, llvm-commits, kristof.beyls
Differential Revision: https://reviews.llvm.org/D45541
llvm-svn: 331601
The main goal is to share getMatchTable between the Instruction
Selector and the Testgen.
The commit also contains some NFC only loosely related to refactoring
out the getMatchTable, but strongly related to the initial Testgen
patch (see https://reviews.llvm.org/D43962)
Reviewers: dsanders, aemerson
Reviewed By: dsanders
Subscribers: rovka, kristof.beyls, llvm-commits, dsanders
Differential Revision: https://reviews.llvm.org/D46096
llvm-svn: 331395
This patch changes GlobalISelEmitter to rank patterns similar to how the
DAG does it (ie it computes a score for a pattern and adds the added
complexity to it).
This is so that the decision tree for GISelSelector remains compatible
with that of SelectionDAG.
https://reviews.llvm.org/D43270
llvm-svn: 325401
Summary:
Apparently, we missed on constraining register classes of VReg-operands of all the instructions
built from a destination pattern but the root (top-level) one. The issue exposed itself
while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped
into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained,
while nested VTOSIZS (or rather its destination virtual register to be exact) does not.
Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI.
https://bugs.llvm.org/show_bug.cgi?id=35965
rdar://problem/36886530
Patch by Roman Tereshin
Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan
Reviewed By: dsanders, qcolombet, rovka
Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D42565
llvm-svn: 323692
Summary:
This patch adds CustomRenderer which renders the matched
operands to the specified instruction.
Targets can enable the matching of SDNodeXForm by adding
a definition that inherits from GICustomOperandRenderer and
GISDNodeXFormEquiv as follows.
def gi_imm8 : GICustomOperandRenderer<"renderImm8”>,
GISDNodeXFormEquiv<imm8_xform>;
Custom renderer functions should be of the form:
void render(MachineInstrBuilder &MIB, const MachineInstr &I);
Reviewers: dsanders, ab, rovka
Reviewed By: dsanders
Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet
Differential Revision: https://reviews.llvm.org/D42012
llvm-svn: 322582
NFC for currently supported targets. This resolves a problem encountered by
targets such as RISCV that reference `Subtarget` in ImmLeaf predicates.
llvm-svn: 321176
*** Context ***
Prior to this patchw, the table generated for matching instruction was
straight forward but highly inefficient.
Basically, each pattern generates its own set of self contained checks
and actions.
E.g., TableGen generated:
// First pattern
CheckNumOperand 3
CheckOpcode G_ADD
...
Build ADDrr
// Second pattern
CheckNumOperand 3
CheckOpcode G_ADD
...
Build ADDri
// Third pattern
CheckNumOperand 3
CheckOpcode G_SUB
...
Build SUBrr
*** Problem ***
Because of that generation, a *lot* of check were redundant between each
pattern and were checked every single time until we reach the pattern
that matches.
E.g., Taking the previous table, let say we are matching a G_SUB, that
means we were going to check all the rules for G_ADD before looking at
the G_SUB rule. In particular we are going to do:
check 3 operands; PASS
check G_ADD; FAIL
; Next rule
check 3 operands; PASS (but we already knew that!)
check G_ADD; FAIL (well it is still not true)
; Next rule
check 3 operands; PASS (really!!)
check G_SUB; PASS (at last :P)
*** Proposed Solution ***
This patch introduces a concept of group of rules (GroupMatcher) that
share some predicates and only get checked once for the whole group.
This patch only creates groups with one nesting level. Conceptually
there is nothing preventing us for having deeper nest level. However,
the current implementation is not smart enough to share the recording
(aka capturing) of values. That limits its ability to do more sharing.
For the given example the current patch will generate:
// First group
CheckOpcode G_ADD
// First pattern
CheckNumOperand 3
...
Build ADDrr
// Second pattern
CheckNumOperand 3
...
Build ADDri
// Second group
CheckOpcode G_SUB
// Third pattern
CheckNumOperand 3
...
Build SUBrr
But if we allowed several nesting level, it could create a sub group
for the checknumoperand 3.
(We would need to call optimizeRules on the rules within a group.)
*** Result ***
With only one level of nesting, the instruction selection pass is up
to 4x faster. For instance, one instruction now takes 500 checks,
instead of 24k! With more nesting we could get in the tens I believe.
Differential Revision: https://reviews.llvm.org/D39034
rdar://problem/34670699
llvm-svn: 321017
This patch splits atomics out of the generic G_LOAD/G_STORE and into their own
G_ATOMIC_LOAD/G_ATOMIC_STORE. This is a pragmatic decision rather than a
necessary one. Atomic load/store has little in implementation in common with
non-atomic load/store. They tend to be handled very differently throughout the
backend. It also has the nice side-effect of slightly improving the common-case
performance at ISel since there's no longer a need for an atomicity check in the
matcher table.
All targets have been updated to remove the atomic load/store check from the
G_LOAD/G_STORE path. AArch64 has also been updated to mark
G_ATOMIC_LOAD/G_ATOMIC_STORE legal.
There is one issue with this patch though which also affects the extending loads
and truncating stores. The rules only match when an appropriate G_ANYEXT is
present in the MIR. For example,
(G_ATOMIC_STORE (G_TRUNC:s16 (G_ANYEXT:s32 (G_ATOMIC_LOAD:s16 X))))
will match but:
(G_ATOMIC_STORE (G_ATOMIC_LOAD:s16 X))
will not. This shouldn't be a problem at the moment, but as we get better at
eliminating extends/truncates we'll likely start failing to match in some
cases. The current plan is to fix this in a patch that changes the
representation of extending-load/truncating-store to allow the MMO to describe
a different type to the operation.
llvm-svn: 319691
GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider
range of orderings. This has then been used to import patterns using nodes such
as atomic_cmp_swap, atomic_swap, and atomic_load_*.
llvm-svn: 319232
Summary:
This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV,
causes TableGen to instrument the generated table to collect rule coverage
information. However, LLVM_ENABLE_GISEL_COV goes a bit further than
LLVM_ENABLE_DAGISEL_COV. The information is written to files
(${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be
concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will
read this information and use it to emit warnings about untested rules.
This technique could also be used by SelectionDAG and can be further
extended to detect hot rules and give them priority over colder rules.
Usage:
* Enable LLVM_ENABLE_GISEL_COV in CMake
* Build the compiler and run some tests
* cat gisel-coverage-[0-9]* > gisel-coverage-all
* Delete lib/Target/*/*GenGlobalISel.inc*
* Build the compiler
Known issues:
* ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual
step due to a lack of a portable 'cat' command. It should be the
concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files.
* There's no mechanism to discard coverage information when the ruleset
changes
Depends on D39742
Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka
Reviewed By: rovka
Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits
Differential Revision: https://reviews.llvm.org/D39747
llvm-svn: 318356
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
