This reverts commit f9599bbc7a3f831e1793a549d8a7a19265f3e504.
For some reason it caused us a huge compile time regression in downstream
workloads. Not sure whether the source of it is in upstream code ir not.
Temporarily reverting until investigated.
Differential Revision: https://reviews.llvm.org/D142330
[Originally committed as f6ddf7781471b71243fa3c3ae7c93073f95c7dff;
reverted in bbef38352fbade9e014ec97d5991da5dee306da7 due to test
breakage; now relanded with the Arm tests conditioned on
`arm-registered-target`]
The LowerTypeTests pass emits a jump table in the form of an
`inlineasm` IR node containing a string representation of some
assembly. It tests the target triple to see what architecture it
should be generating assembly for. But that's not good enough for
`Triple::thumb`, because the 32-bit PC-relative `b.w` branch
instruction isn't available in all supported architecture versions. In
particular, Armv6-M doesn't support that instruction (although the
similar Armv8-M Baseline does).
Most of this patch is concerned with working out whether the
compilation target is Armv6-M or not, which I'm doing by going through
all the functions in the module, retrieving a TargetTransformInfo for
each one, and querying it via a new method I've added to check its
SubtargetInfo. If any function's TTI indicates that it's targeting an
architecture supporting B.W, then we assume we're also allowed to use
B.W in the jump table.
The Armv6-M compatible jump table format requires a temporary
register, and therefore also has to use the stack in order to restore
that register.
Another consequence of this change is that jump tables on Arm/Thumb
are no longer always the same size. In particular, on an architecture
that supports Arm and Thumb-1 but not Thumb-2, the Arm and Thumb
tables are different sizes from //each other//. As a consequence,
``getJumpTableEntrySize`` can no longer base its answer on the target
triple's architecture: it has to take into account the decision that
``selectJumpTableArmEncoding`` made, which meant I had to move that
function to an earlier point in the code and store its answer in the
``LowerTypeTestsModule`` class.
Reviewed By: lenary
Differential Revision: https://reviews.llvm.org/D143576
InstCombine is supposed to be a superset of InstSimplify, but
failed to invoke load simplification.
Unfortunately, this causes a minor compile-time regression, which
will be mitigated in a future commit.
define i1 @compare_vscales() {
%vscale = call i64 @llvm.vscale.i64()
%vscalex2 = shl nuw nsw i64 %vscale, 1
%vscalex4 = shl nuw nsw i64 %vscale, 2
%cmp = icmp ult i64 %vscalex2, %vscalex4
ret i1 %cmp
}
This IR is currently emitted by LLVM. This icmp is redundant as this snippet
can be simplified to true or false as both operands originate from the same
@llvm.vscale.i64() call.
Differential Revision: https://reviews.llvm.org/D142542
I stumbled over this while trying to improve our exit count work. These expressions
are equivalent for complementary signed/unsigned ext and min/max (including umin_seq),
but they are not canonicalized and SCEV cannot recognize them as the same.
The benefit of this canonicalization is that SCEV can prove some new equivalences which
it coudln't prove because of different forms. There is 1 test where trip count seems pessimized,
I could not directly figure out why, but it just seems an unrelated issue that we can fix.
Other changes seem neutral or positive to me.
Differential Revision: https://reviews.llvm.org/D141481
Reviewed By: nikic
We were not summarizing a function alias in the vtable, leading to
incorrect WPD in some cases, and missing WPD in others.
Specifically, we would end up ignoring function aliases as they aren't
summarized, so we could incorrectly devirtualize if there was a single
other non-alias function in a compatible vtable. And if there was only
one implementation, but it was an alias, we would not be able to
identify and perform the single implementation devirtualization.
Handling the alias summary correctly also required fixing the handling
in mustBeUnreachableFunction, so that it is not incorrectly ignored.
Regular LTO is conservatively correct because it will skip
devirtualizing when any pointer within a vtable is not a function.
However, it needs additional work to be able to take advantage of
function alias within the vtable that is in fact the only
implementation. For that reason, the Regular LTO testing in the second
test case is currently disabled, and will be enabled along with a follow
on enhancement fix for Regular LTO WPD.
Differential Revision: https://reviews.llvm.org/D144209
This reverts commit f6ddf7781471b71243fa3c3ae7c93073f95c7dff.
Eight buildbots reported that the two test files changed by that
commit had started failing. The buildbots in question all had in
common that they build with a very restricted `LLVM_TARGETS_TO_BUILD`,
such as only X86 or AArch64 or Hexagon. I didn't notice this before
commit because my own build has the full default set of targets, and
in that circumstance, the tests pass.
I assume the problem has something to do with the attempt to query
TargetTransformInfo: if you can't make a valid TTI for the target
triple then you can't ask it what kind of inline assembler you should
be emitting, and so `opt` without the Arm backend can't get the Arm
cases of these tests right.
I don't have time to fix this until next week, so I'll revert the
change for now to keep the buildbots happy.
The LowerTypeTests pass emits a jump table in the form of an
`inlineasm` IR node containing a string representation of some
assembly. It tests the target triple to see what architecture it
should be generating assembly for. But that's not good enough for
`Triple::thumb`, because the 32-bit PC-relative `b.w` branch
instruction isn't available in all supported architecture versions. In
particular, Armv6-M doesn't support that instruction (although the
similar Armv8-M Baseline does).
Most of this patch is concerned with working out whether the
compilation target is Armv6-M or not, which I'm doing by going through
all the functions in the module, retrieving a TargetTransformInfo for
each one, and querying it via a new method I've added to check its
SubtargetInfo. If any function's TTI indicates that it's targeting an
architecture supporting B.W, then we assume we're also allowed to use
B.W in the jump table.
The Armv6-M compatible jump table format requires a temporary
register, and therefore also has to use the stack in order to restore
that register.
Another consequence of this change is that jump tables on Arm/Thumb
are no longer always the same size. In particular, on an architecture
that supports Arm and Thumb-1 but not Thumb-2, the Arm and Thumb
tables are different sizes from //each other//. As a consequence,
``getJumpTableEntrySize`` can no longer base its answer on the target
triple's architecture: it has to take into account the decision that
``selectJumpTableArmEncoding`` made, which meant I had to move that
function to an earlier point in the code and store its answer in the
``LowerTypeTestsModule`` class.
Reviewed By: lenary
Differential Revision: https://reviews.llvm.org/D143576
Update ConstraintSystem to use a sparse representation for entries in a
row. Most rows only contain a small number of variables, so the sparse
representation can result in significant speedups.
For a large test case from D135915, it halves the time spent in
ConstraintElimination.
To ensure this returns the same results as the old implementation in all
cases, I built a large set of projects with an extra assertion that it
produces the same result as the old implementation.
We can only fold insertvalue undef, (extractvalue x, n) to x
if x is not poison, otherwise we might be replacing undef with
poison (https://alive2.llvm.org/ce/z/fnw3c8). The insertvalue
poison case is always fine.
I didn't go to particularly large effort to preserve cases where
folding with undef is still legal (mainly when there is a chain of
multiple inserts that end up covering the whole aggregate),
because this shouldn't really occur in practice: We should always
be generating the insertvalue poison form when constructing
aggregates nowadays.
Differential Revision: https://reviews.llvm.org/D144106
This reverts commit 2a2a6bfcfe8e62886542cb673ac8df349cf26499.
This causes build failures:
/home/npopov/repos/llvm-project/llvm/lib/Analysis/ConstraintSystem.cpp: In member function ‘llvm::SmallVector<std::__cxx11::basic_string<char> > llvm::ConstraintSystem::getVarNamesList() const’:
/home/npopov/repos/llvm-project/llvm/lib/Analysis/ConstraintSystem.cpp:118:10: error: invalid use of incomplete type ‘class llvm::Value’
118 | if (V->getName().empty())
| ^~
In file included from /home/npopov/repos/llvm-project/llvm/lib/Analysis/ConstraintSystem.cpp:9:
/home/npopov/repos/llvm-project/llvm/include/llvm/Analysis/ConstraintSystem.h:21:7: note: forward declaration of ‘class llvm::Value’
21 | class Value;
| ^~~~~
/home/npopov/repos/llvm-project/llvm/lib/Analysis/ConstraintSystem.cpp:119:22: error: invalid use of incomplete type ‘class llvm::Value’
119 | OperandName = V->getNameOrAsOperand();
| ^~
/home/npopov/repos/llvm-project/llvm/include/llvm/Analysis/ConstraintSystem.h:21:7: note: forward declaration of ‘class llvm::Value’
21 | class Value;
| ^~~~~
/home/npopov/repos/llvm-project/llvm/lib/Analysis/ConstraintSystem.cpp:121:41: error: invalid use of incomplete type ‘class llvm::Value’
121 | OperandName = std::string("%") + V->getName().str();
| ^~
/home/npopov/repos/llvm-project/llvm/include/llvm/Analysis/ConstraintSystem.h:21:7: note: forward declaration of ‘class llvm::Value’
21 | class Value;
| ^~~~~
Previously when constraint system outputs the rows in the system the variables used are x1,2...n making it hard to infer which ones they relate to in the IR
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D142618
Check that the underlying object is a constant global with
definitive initializer upfront, so we can skip the more expensive
offset calculation logic if we can't perform the fold anyway.
There are 2 issues here:
1. In the default LLVM FP environment (regular FP math instructions),
SNaN is some flavor of "don't care" which we will nail down in
D143074, so this is just a quality-of-implementation improvement
for default FP.
2. In the constrained FP environment (constrained intrinsics), SNaN
must not propagate through a math operation; it has to be quieted
according to IEEE-754 spec. That is independent of exception
handling mode, so the current behavior is a miscompile.
Differential Revision: https://reviews.llvm.org/D143505
This fixes the handling of phi nodes in getConstantRangeAtUse()
and re-enables it, reverting the workaround from
c77c186a647b385c291ddabecd70a2b4f84ae342.
For phi nodes, while we can make use of the edge condition for the
incoming value, we shouldn't look past the phi node to look for
further conditions, because we might be reasoning about values
from two different cycle iterations (which will have the same
SSA value).
To handle this more specifically we would have to detect cycles,
and there doesn't seem to be any motivating case for that at this
point.
This effectively reverts 5c38c6a and 4f772b0.
A recently introduced LazyValueInfo::getConstantRangeAtUse returns incorrect
ranges for values in certain cases. One such example is described in PR60629.
The issue has something to do with traversing PHI uses of a value transitively.
As nikic pointed out, we're effectively reasoning about values from different
loop iterations.
In the faulting test case, CVP made a miscompilation because the calculated
range for a shift argument was incorrect. It returned empty-set, however it is
clearly not a dead code. CVP then erased the shift instruction because
of empty range.
The one caller of this function already checked that V isn't a
Constant.
Alternatively, we could remove the check from the caller if reviewers
prefer.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D143677
I guess its only reason to exist is potential CT optimization, otherwise it is
just creating cohesion between this code and rewriter internals. We plan to
extend the rewriter. I'd rather not have this cohesion, unless there is a serious
reason to have it.
Differential Revision: https://reviews.llvm.org/D143246
There are some helpers in the Lint analysis pass that will setup
a pass manager and then run the Lint pass on a given Function/Module.
Those have been using the LegacyPassManager, but as a small step
towards removing the deprecated legacy pass manager this patch is
changing those helpers into using the new pass manager instead.
No idea if anyone is really is using those helpers. Maybe an
alternative had been to just remove them. There is at least no unit
tests or similar that verifies that they work, so I validated this
patch by using a hacked opt binary that called those functions
before running the normal pipeline.
Differential Revision: https://reviews.llvm.org/D143388
This patch breaks up the solving step into 2 phases:
1. Collect all rows where the variable to eliminate is != 0 and remove
it from the original system.
2. Process all collect rows to build new set of constraints, add them to
the original system.
This is much more efficient for excessive cases, as this avoids a large
number of moves to the new system. This reduces the time spent in
ConstraintElimination for the test case shared in D135915 from ~3s to
0.6s.
This reverts commit a772f0bb920a4957fb94dd8dbe45943809fd0ec3.
The main problem was related to how we handled `dbgs()` from the hosted
compiler. Using explicit `subprocess.communicate`, and not relying on
dbgs() being flushed until the end appears to address the problem.
Also some fixes due to some bots running older pythons, so we can't have
nice things like `int | float` and such.
This reverts commit a7354899d1a235a796b3a2ccb45f6596983c8672.
The way stdout/stderr get routed seems to work differently locally and
on the bots. Investigating.
This hooks up the interactive model runner to the passes that support
ml-based decisions. Because the interface to this runner is the exact
same as the one used during inference, we just reuse the exact same
setup we have for "release mode". This makes "release mode" a misnomer -
and that's something we needed to resolve sooner or later (e.g.
supporting more than one embedded model for the same problem was another
reason to drop that nomenclature). That will happen in a subsequent
change.
To use this evaluator, just enable the pass in (currently) "release"
mode, but also pass the base name for the 2 channel files via the
pass-specific flag.
The 2 files are the responsibilty of the hosting process. The added
tests use a minimal, toy such host, illustrating setup and
communication.
Differential Revision: https://reviews.llvm.org/D143218
Delete the opt intrinsics since they are now identical.
I left the side effects due to user expectations about how these
interact with things like inline assembly or function calls. Or
that they wouldn't be hoisted. I think we should look at other
ways to address thoughs.
If I could, I'd rename them these somehow to distance them from
the vsetvli instruction. In some sense they only query the VL for
a particular SEW and LMUL. They don't guarantee a vsetvli
instruction will be emitted.
Fixes https://github.com/llvm/llvm-project/issues/59359
Reviewed By: rogfer01, kito-cheng
Differential Revision: https://reviews.llvm.org/D143220
This NFC (intended) patch has several small changes:
* It renames PredicationStyle to TailFoldingStyle.
* It renames TTI.emitActiveLaneMask() to TTI.getPreferredTailFoldingStyle()
* Simplifies some of its uses in the LoopVectorizer
Rationale: To my surprise PredicationStyle::None did not mean 'no
predication', but rather 'no active lane mask intrinsic', such that the
predicate is created using a splat + compare with stepvector. The enum is
also highly specific to tail folding, so it seems better to name this
around that feature, i.e. 'tail folding style'.
This also makes it more amenable to extend it to other tail folding styles,
such as the one added in D142109.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D142887
