createCast in MergeFunctions did not consider ArrayTypes, which results
in the creation of a bitcast between ArrayTypes in the thunk function,
leading to an assertion failure in the provided test case.
The version of createCast in GlobalMergeFunctions does handle
ArrayTypes, so this common code has been factored out into the
IRBuilder.
This reapplies #132522.
Previously casts of scalable m_ImmConstant splats weren't being folded
by ConstantFoldCastOperand, triggering the "Constant-fold of ImmConstant
should not fail" assertion.
There are no changes to the code in this PR, instead we just needed
#133207 to land first.
A test has been added for the assertion in
llvm/test/Transforms/InstSimplify/vec-icmp-of-cast.ll
@icmp_ult_sext_scalable_splat_is_true.
<hr/>
#118806 fixed an infinite loop in FoldShiftByConstant that could occur
when the shift amount was a ConstantExpr.
However this meant that FoldShiftByConstant no longer kicked in for
scalable vectors because scalable splats are represented by
ConstantExprs.
This fixes it by allowing scalable splats of non-ConstantExprs in
m_ImmConstant, which also fixes a few other test cases where scalable
splats were being missed.
But I'm also hoping that UseConstantIntForScalableSplat will eventually
remove the need for this.
I noticed this when trying to reverse a combine on RISC-V in #132245,
and saw that the resulting vector and scalar forms were different.
Previously only fixed vector splats were handled. This adds supports for
scalable vectors too by allowing ConstantExpr splats.
We need to add the extra V->getType()->isVectorTy() check because a
ConstantExpr might be a scalar to vector bitcast.
By allowing ConstantExprs this also allow fixed vector ConstantExprs to
be folded, which causes the diffs in
llvm/test/Analysis/ValueTracking/known-bits-from-operator-constexpr.ll
and llvm/test/Transforms/InstSimplify/ConstProp/cast-vector.ll. I can
remove them from this PR if reviewers would prefer.
Fixes#132922
In the profitability check for vectorization, the dependency matrix was
not handled correctly. This can result to make a wrong decision: It may
say "this loop can be vectorized" when in fact it cannot. The root cause
of this is that the check process early returns when it finds '=' or 'I'
in the dependency matrix. To make sure that we can actually vectorize
the loop, we need to check all the rows of the matrix. This patch fixes
the process of checking whether we can vectorize the loop or not. Now it
won't make a wrong decision for a loop that cannot be vectorized.
Related: #131130
#118806 fixed an infinite loop in FoldShiftByConstant that could occur
when the shift amount was a ConstantExpr.
However this meant that FoldShiftByConstant no longer kicked in for
scalable vectors because scalable splats are represented by
ConstantExprs.
This fixes it by allowing scalable splats of non-ConstantExprs in
m_ImmConstant, which also fixes a few other test cases where scalable
splats were being missed.
But I'm also hoping that UseConstantIntForScalableSplat will eventually
remove the need for this.
I noticed this when trying to reverse a combine on RISC-V in #132245,
and saw that the resulting vector and scalar forms were different.
---------
Co-authored-by: Yingwei Zheng <dtcxzyw@qq.com>
Currently iterators over EquivalenceClasses will iterate over std::set,
which guarantees the order specified by the comperator. Unfortunately in
many cases, EquivalenceClasses are used with pointers, so iterating over
std::set of pointers will not be deterministic across runs.
There are multiple places that explicitly try to sort the equivalence
classes before using them to try to get a deterministic order
(LowerTypeTests, SplitModule), but there are others that do not at the
moment and this can result at least in non-determinstic value naming in
Float2Int.
This patch updates EquivalenceClasses to keep track of all members via a
extra SmallVector and removes code from LowerTypeTests and SplitModule
to sort the classes before processing.
Overall it looks like compile-time slightly decreases in most cases, but
close to noise:
https://llvm-compile-time-tracker.com/compare.php?from=7d441d9892295a6eb8aaf481e1715f039f6f224f&to=b0c2ac67a88d3ef86987e2f82115ea0170675a17&stat=instructions
PR: https://github.com/llvm/llvm-project/pull/134075
Need to update the mapping between gathered values and their matching
entries, if the list of the entries is updated and only some of them are
selected for final shuffling.
Fixes#134085
Preserve branch weight metadata when merging instructions if one of the
instructions is missing metadata. This is similar in behaviour to what
we do today for other types of metadata such as mmra, memprof and
callsite metadata.
The test suite of LoopVectorize suffers from a coverage hole when types
mismatch, and runtime checks are needed, with a conflict redux. Fix this
coverage hole by adding tests.
There is a problem with the current profitability check for
vectorization in LoopInterchange. There are both false positives and
false negatives. The former means that the heuristic may say that "an
exchange is necessary to vectorize the innermost loop" even though it's
already possible. The latter means that the heuristic may miss a case
where an exchange is necessary to vectorize the innermost loop. Note
that this is not a dependency analysis problem. This is caused by
incorrect handling of the dependency matrix in the profitability check,
so these problems can occur even if the analysis is accurate (no
overestimation).
This patch adds tests to clarify the cases that should be fixed. The
root cause of these cases is that the heuristic doesn't handle the
direction of a dependency correctly.
As noted in 1a9358c090d0507be21c5e9b2d97a23ef1de8ab0, some
simplifications can produce a redundant select where the true and false
operands are the same, which this patch removes.
The is_fpclass test was changed so the condition wasn't made dead.
LoopInterchange has several heuristic functions to determine if
exchanging two loops is profitable or not. Whether or not to use each
heuristic and the order in which to use them were fixed, but #125830
allows them to be changed internally at will. This patch adds a new
option to control them via the compiler option.
The previous patch also added an option to prioritize the vectorization
heuristic. This patch also removes it to avoid conflicts between it and
the newly introduced one, e.g., both
`-loop-interchange-prioritize-vectorization=1` and
`-loop-interchange-profitabilities='cache,vectorization'` are specified.
Vectorizing of fminimumnum and fminimumnum have not support yet. Let's
add the testcase for it now, and we will update the testcase when we
support it.
If the scalar instructions is marked for the vectorization in the tree,
it cannot be vectorized as part of the another node in the same tree, in
general. It may prevent some potentially profitable vectorization
opportunities, since some nodes end up being buildvector/gather nodes,
which add to the total cost.
Patch allows revectorization of the previously vectorized scalars.
Reviewers: hiraditya, RKSimon
Reviewed By: RKSimon, hiraditya
Pull Request: https://github.com/llvm/llvm-project/pull/133091
Not sure why the "fold-all" option naming didn't match the
variable "FoldPreOutputs", but I've preserved the difference.
More annoyingly, the pass name "normalize" does not match the pass
name IRNormalizer and should probably be fixed one way or the other.
Also the existing test coverage for the flags is lacking. I've added
a test that shows they parse, but we should have tests that they
do something.
During the transition from debug intrinsics to debug records, we used
several different command line options to customise handling: the
printing of debug records to bitcode and textual could be independent of
how the debug-info was represented inside a module, whether the
autoupgrader ran could be customised. This was all valuable during
development, but now that totally removing debug intrinsics is coming
up, this patch removes those options in favour of a single flag
(experimental-debuginfo-iterators), which enables autoupgrade, in-memory
debug records, and debug record printing to bitcode and textual IR.
We need to do this ahead of removing the
experimental-debuginfo-iterators flag, to reduce the amount of
test-juggling that happens at that time.
There are quite a number of weird test behaviours related to this --
some of which I simply delete in this commit. Things like
print-non-instruction-debug-info.ll , the test suite now checks for
debug records in all tests, and we don't want to check we can print as
intrinsics. Or the update_test_checks tests -- these are duplicated with
write-experimental-debuginfo=false to ensure file writing for intrinsics
is correct, but that's something we're imminently going to delete.
A short survey of curious test changes:
* free-intrinsics.ll: we don't need to test that debug-info is a zero
cost intrinsic, because we won't be using intrinsics in the future.
* undef-dbg-val.ll: apparently we pinned this to non-RemoveDIs in-memory
mode while we sorted something out; it works now either way.
* salvage-cast-debug-info.ll: was testing intrinsics-in-memory get
salvaged, isn't necessary now
* localize-constexpr-debuginfo.ll: was producing "dead metadata"
intrinsics for optimised-out variable values, dbg-records takes the
(correct) representation of poison/undef as an operand. Looks like we
didn't update this in the past to avoid spurious test differences.
* Transforms/Scalarizer/dbginfo.ll: this test was explicitly testing
that debug-info affected codegen, and we deferred updating the tests
until now. This is just one of those silent gnochange issues that get
fixed by RemoveDIs.
Finally: I've added a bitcode test, dbg-intrinsics-autoupgrade.ll.bc,
that checks we can autoupgrade debug intrinsics that are in bitcode into
the new debug records.
Vectorizing of fminimumnum and fminimumnum have not support yet. Let's
add the testcase for it now, and we will update the testcase when we
support it.
getSameOpcode in some cases may consider 2 compares as having same
opcode, even though previously they were considered as alternate. It may
happen, because getSameOpcode looses info about previous instructions
and their states. Need to use isAlternateInstruction function instead
for the correct analysis.
Reviewers: RKSimon, hiraditya
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/133769
If a recipe was predicated and tail folded at the same time, it will
have a mask like
EMIT vp<%header-mask> = icmp ule canonical-iv, backedge-tc
EMIT vp<%mask> = logical-and vp<%header-mask>, vp<%pred-mask>
When converting to an EVL recipe, if the mask isn't exactly just the
header-mask we copy the whole logical-and.
We can remove this redundant logical-and (because it's now covered by
EVL) and just use vp<%pred-mask> instead.
This lets us remove the widened canonical IV in more places.
Need to check the value type, not the return type, of the instructions,
when doing the analysis for the whole register use to prevent a compiler
crash.
Fixes#133751
I recently added a new option to update_test_checks.py that can
filter out all CHECK lines after a certain point. We usually don't
care about checking for the original scalar loop after the vector
loop because it doesn't change. Cutting out unnecessary CHECK
lines makes the files smaller and hopefully the tests run quicker.
Remove the UF = 1 restriction introduced by 577631f0a5 building on top
of 783a846507683, which allows updating all relevant users of the VF,
VPScalarIVSteps in particular.
This restores the full functionality of
https://github.com/llvm/llvm-project/pull/106441.
If UnJ completely unrolls a loop and removes it entirely, the loop
remains in the current loop nest. If the loop nest gets reused the loops
will no longer be valid. As there is no way to remove a loop from a
LoopNest, this patch removes the preserve of the LoopNestAnalysis so
that it will be regenerated.
Fixes#124518
