Summary:
This is a more general fix to 59029b9eef (D75704).
This patch does the following:
1. updates isKnownBaseValue to account for base pointer and
derived pointer having differing types.
2. This inturn allows us to populate the
lattice (States) for such derived pointers.
3. It also updates all states where the base and derived pointers have
differing types (vector versus scalar) and conservatively marks these
states as conflictcs.
Note that in 59029b9eef, we were just fixing existing lattice values
and that too, only for uses of extractelement.
Reviewers: reames, skatkov, dantrushin
Reviewed By: skatkov
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76305
Expands on the enablement of the shouldSinkOperands() TLI hook in:
D79718
The last codegen/IR test diff shows what I suspected could happen - we were
sinking all splat shift operands into a loop. But that's not what we want in
general; we only want to sink the *shift amount* operand if it is a splat.
Differential Revision: https://reviews.llvm.org/D79827
This patch introduces an improvement in the Alignment of the loads
generated in createReplacementValues() by querying AAAlign attribute for
the best Alignment for the base.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D76550
Summary:
This change introduces InliningAdvisor (and related APIs), the interface
that abstracts decision making away from the inlining pass. We will use
this interface to delegate decision making to a trained ML model,
subsequently (see referenced RFC).
RFC: http://lists.llvm.org/pipermail/llvm-dev/2020-April/140763.html
Reviewers: davidxl, eraman, dblaikie
Subscribers: mgorny, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79042
Summary:
Update check to include the check for unreachable.
Basic blocks ending in unreachable are special cased, as these blocks may be already unswitched. Before this patch this check is only done for the default destination.
The condition for the exit cases and the default case must be the same, because we should never leave edges from the switch instruction to a basic block that we are unswitching. In PR45355 we still have a remaining edge (that we're attempting to remove from the DT) because its the default edge to an unreachable-terminated block where we unswitch a case edge to that block.
Resolves PR45355.
Reviewers: chandlerc
Subscribers: hiraditya, uabelho, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78279
This patch adds a new TTI hook to allow targets to tell LSR that
a chain including some instruction is already profitable and
should not be optimized. This patch also adds an implementation
of this TTI hook for ARM so LSR doesn't optimize chains that include
the VCTP intrinsic.
Differential Revision: https://reviews.llvm.org/D79418
This was reverted because of a miscompilation. At closer inspection, the
problem was actually visible in a changed llvm regression test too. This
one-line follow up fix/recommit will splat the IV, which is what we are trying
to avoid if unnecessary in general, if tail-folding is requested even if all
users are scalar instructions after vectorisation. Because with tail-folding,
the splat IV will be used by the predicate of the masked loads/stores
instructions. The previous version omitted this, which caused the
miscompilation. The original commit message was:
If tail-folding of the scalar remainder loop is applied, the primary induction
variable is splat to a vector and used by the masked load/store vector
instructions, thus the IV does not remain scalar. Because we now mark
that the IV does not remain scalar for these cases, we don't emit the vector IV
if it is not used. Thus, the vectoriser produces less dead code.
Thanks to Ayal Zaks for the direction how to fix this.
This is a reimplementation of the `orderNodes` function, as the old
implementation didn't take into account all cases.
Fix PR41509
Differential Revision: https://reviews.llvm.org/D79037
Fix the assumption that all bitcasts of the same type sizes are free.
We now only assume that bitcasts between ints and ptrs of the same
size are free. This allows TTImpl to just call the concrete
implementation of getCastInstrCost.
Differential Revision: https://reviews.llvm.org/D78918
Fixes PR41696
The loop-reroll pass generates an invalid IR (or its assertion
fails in debug build) if values of the base instruction and
other root instructions (terms used in the loop-reroll pass)
are used outside the loop block. See IRs written in PR41696
as examples.
The current implementation of the loop-reroll pass can reroll
only loops that don't have values that are used outside the
loop, except reduced values (the last values of reduction chains).
This is described in the comment of the `LoopReroll::reroll`
function.
https://github.com/llvm/llvm-project/blob/llvmorg-10.0.0/llvm/lib/Transforms/Scalar/LoopRerollPass.cpp#L1600
This is checked in the `LoopReroll::DAGRootTracker::validate`
function.
https://github.com/llvm/llvm-project/blob/llvmorg-10.0.0/llvm/lib/Transforms/Scalar/LoopRerollPass.cpp#L1393
However, the base instruction and other root instructions skip
this check in the validation loop.
https://github.com/llvm/llvm-project/blob/llvmorg-10.0.0/llvm/lib/Transforms/Scalar/LoopRerollPass.cpp#L1229
Moving the check in front of the skip is the logically simplest
fix. However, inserting the check in an earlier stage is better
in terms of compilation time of unrerollable loops. This fix
inserts the check for the base instruction into the function
to validate possible base/root instructions. Check for other
root instructions is unnecessary because they don't match any
base instructions if they have uses outside the loop.
Differential Revision: https://reviews.llvm.org/D79549
For AAReturnedValues we treated new and existing information differently
in the updateImpl. Only the latter was properly analyzed and
categorized. The former was thought to be analyzed in the subsequent
update. Since the Attributor does not support "self-updates" we need to
make sure the state is "stable" after each updateImpl invocation. That
is, if the surrounding information does not change, the state is valid.
Now we make sure all return values have been handled and properly
categorized each iteration. We might not update again if we have not
requested a non-fix attribute so we cannot "wait" for the next update to
analyze a new return value.
Bug reported by @sdmitriev.
Summary:
This fixes PR45885 by fixing isGuaranteedNotToBeUndefOrPoison so it does not look into dominating
branch conditions of V when V is an instruction in an unreachable block.
Reviewers: spatel, nikic, lebedev.ri
Reviewed By: nikic
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79790
We want to add a way to avoid merging identical calls so as to keep the
separate debug-information for those calls. There is also an asan
usecase where having this attribute would be beneficial to avoid
alternative work-arounds.
Here is the link to the feature request:
https://bugs.llvm.org/show_bug.cgi?id=42783.
`nomerge` is different from `noline`. `noinline` prevents function from
inlining at callsites, but `nomerge` prevents multiple identical calls
from being merged into one.
This patch adds `nomerge` to disable the optimization in IR level. A
followup patch will be needed to let backend understands `nomerge` and
avoid tail merge at backend.
Reviewed By: asbirlea, rnk
Differential Revision: https://reviews.llvm.org/D78659
SDAG suffers when it can't see that a funnel operand is a splat value
(due to single-basic-block visibility), so invert the normal loop
hoisting rules to move a splat op closer to its use.
This would be part 1 of an enhancement similar to D63233.
This is needed to re-fix PR37426:
https://bugs.llvm.org/show_bug.cgi?id=37426
...because we got better at canonicalizing IR to funnel shift intrinsics.
The existing CGP code for shift opcodes is likely overstepping what it was
intended to do, so that will be fixed in a follow-up.
Differential Revision: https://reviews.llvm.org/D79718
GNU ld's internal linker script uses (https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=add44f8d5c5c05e08b11e033127a744d61c26aee)
.text :
{
*(.text.unlikely .text.*_unlikely .text.unlikely.*)
*(.text.exit .text.exit.*)
*(.text.startup .text.startup.*)
*(.text.hot .text.hot.*)
*(SORT(.text.sorted.*))
*(.text .stub .text.* .gnu.linkonce.t.*)
/* .gnu.warning sections are handled specially by elf.em. */
*(.gnu.warning)
}
Because `*(.text.exit .text.exit.*)` is ordered before `*(.text .text.*)`, in a -ffunction-sections build, the C library function `exit` will be placed before other functions.
gold's `-z keep-text-section-prefix` has the same problem.
In lld, `-z keep-text-section-prefix` recognizes `.text.{exit,hot,startup,unlikely,unknown}.*`, but not `.text.{exit,hot,startup,unlikely,unknown}`, to avoid the strange placement problem.
In -fno-function-sections or -fno-unique-section-names mode, a function whose `function_section_prefix` is set to `.exit"`
will go to the output section `.text` instead of `.text.exit` when linked by lld.
To address the problem, append a dot to become `.text.exit.`
Reviewed By: grimar
Differential Revision: https://reviews.llvm.org/D79600
gcov 4.8 (r189778) moved the exit block from the last to the second.
The .gcda format is compatible with 4.7 but
* decoding libgcov 4.7 produced .gcda with gcov [4.7,8) can mistake the
exit block, emit bogus `%s:'%s' has arcs from exit block\n` warnings,
and print wrong `" returned %s` for branch statistics (-b).
* decoding libgcov 4.8 produced .gcda with gcov 4.7 has similar issues.
Also, rename "return block" to "exit block" because the latter is the
appropriate term.
- Specifically check for sext/zext users which have 'long' form NEON
instructions.
- Add more entries to the table for sext/zexts so that we can report
more accurately the number of vmovls required for NEON.
- Pass the instruction to the pass implementation.
Differential Revision: https://reviews.llvm.org/D79561
We will now ensure ensure the return type of called function is the type
of all call sites we are going to rewrite. This avoids a problem
partially fixed by D79680. The part that was not covered is a use of
this "weird" casted call site (see `@func3` in `misc_crash.ll`).
misc_crash.ll checks are auto-generated now.
We should never give up on AAIsDead as it guards other AAs from
unreachable code (in which SSA properties are meaningless). We did
however use required dependences on some queries in AAIsDead which
caused us to invalidate AAIsDead if the queried AA got invalidated.
We now use optional dependences instead. The bug that exposed this is
added to the liveness.ll test and other test changes show the impact.
Bug report by @sdmitriev.
During an update of AAIsDead, new instructions become live. If we query
information from them, the result is often just the initial state, e.g.,
for call site `noreturn` and `nounwind`. We will now trigger an update
for cached attributes during the AAIsDead update, though other AAs might
later use the same API.
This patch addresses two issues related to adding inline functions to the import list while recursively going through the profiling data.
1. For callsite samples, only add an inlined function to the import list if it's from outside of the module (i.e. only has a declaration inside the module).
2. For body samples, add each target function to the import list if it's from outside of the module (i.e. only has a declaration inside the module). Previously we were using getSubProgram() to check whether it has dbg info, which is inaccurate. This fix properly add imports and could improve the quality of the pass.
Added a few changes to the test to catch these cases.
Differential Revision: https://reviews.llvm.org/D79379
Summary:
this patch fixe crash/asserts found in the test-suite.
the AssumeptionCache cannot be assumed to have all assumes contrary to what i tought.
prevent generation of information for terminators, because this can create broken IR in transfromation where we insert the new terminator before removing the old one.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79458
don't span their entire scope.
The previous commit (6d1c40c171) is an older version of the test.
Reviewed By: aprantl, vsk
Differential Revision: https://reviews.llvm.org/D79573
The old QuerriedAAs contained two vectors, one for required one for
optional dependences (=queries). We now use a single vector and encode
the kind directly in the pointer.
This reduces memory consumption and makes the connection between
abstract attributes and their dependences clearer.
No functional change is intended, changes in the test are due to
different order in the query map. Neither the order before nor now is in
any way special.
---
Single run of the Attributor module and then CGSCC pass (oldPM)
for SPASS/clause.c (~10k LLVM-IR loc):
Before:
```
calls to allocation functions: 543734 (329735/s)
temporary memory allocations: 105895 (64217/s)
peak heap memory consumption: 19.19MB
peak RSS (including heaptrack overhead): 102.26MB
total memory leaked: 269.10KB
```
After:
```
calls to allocation functions: 513292 (341511/s)
temporary memory allocations: 106028 (70544/s)
peak heap memory consumption: 13.35MB
peak RSS (including heaptrack overhead): 95.64MB
total memory leaked: 269.10KB
```
Difference:
```
calls to allocation functions: -30442 (208506/s)
temporary memory allocations: 133 (-910/s)
peak heap memory consumption: -5.84MB
peak RSS (including heaptrack overhead): 0B
total memory leaked: 0B
```
---
Reviewed By: uenoku
Differential Revision: https://reviews.llvm.org/D78729
When we have an existing `argmemonly` or `inaccessiblememorargmemonly`
we used to "know" that information. However, interprocedural constant
propagation can invalidate these attributes. We now ignore and remove
these attributes for internal functions (which may be affected by IP
constant propagation), if we are deriving new attributes for the
function.
As we replace values with constants interprocedurally, we also need to
do this "look-through" step during the generic value traversal or we
would derive properties from replaced values. While this is often not
problematic, it is when we use the "kind" of a value for reasoning,
e.g., accesses to arguments allow `argmemonly`.
We now use getPointerDereferenceableBytes to determine `nonnull` and
`dereferenceable` facts from the IR. We also use getPointerAlignment in
AAAlign for the same reason. The latter can interfere with callbacks so
we do restrict it to non-function-pointers for now.
Defaulting to -Xclang -coverage-version='407*' makes .gcno/.gcda
compatible with gcov [4.7,8)
In addition, delete clang::CodeGenOptionsBase::CoverageExtraChecksum and GCOVOptions::UseCfgChecksum.
We can infer the information from the version.
With this change, .gcda files produced by `clang --coverage a.o` linked executable can be read by gcov 4.7~7.
We don't need other -Xclang -coverage* options.
There may be a mismatching version warning, though.
(Note, GCC r173147 "split checksum into cfg checksum and line checksum"
made gcov 4.7 incompatible with previous versions.)
rL144865 incorrectly wrote function names for GCOV_TAG_FUNCTION
(this might be part of the reasons the header says
"We emit files in a corrupt version of GCOV's "gcda" file format").
rL176173 and rL177475 realized the problem and introduced -coverage-no-function-names-in-data
to work around the issue. (However, the description is wrong.
libgcov never writes function names, even before GCC 4.2).
In reality, the linker command line has to look like:
clang --coverage -Xclang -coverage-version='407*' -Xclang -coverage-cfg-checksum -Xclang -coverage-no-function-names-in-data
Failing to pass -coverage-no-function-names-in-data can make gcov 4.7~7
either produce wrong results (for one gcov-4.9 program, I see "No executable lines")
or segfault (gcov-7).
(gcov-8 uses an incompatible format.)
This patch deletes -coverage-no-function-names-in-data and the related
function names support from libclang_rt.profile
Summary:
The assume builder was non-deterministic when working on unamed values.
this patch fixes this.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: hiraditya, mgrang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78616
Summary: with this patch the assume salvageKnowledge will not generate assume if all knowledge is already available in an assume with valid context. assume bulider can also in some cases update an existing assume with better information.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78014
We have a transform in the opposite direction only for the x86 MMX type,
Other types are not handled either way before this patch.
The motivating case from PR45748:
https://bugs.llvm.org/show_bug.cgi?id=45748
...is the last test diff. In that example, we are triggering an existing
bitcast transform, so we reduce the number of casts, and that should give
us the ideal x86 codegen.
Differential Revision: https://reviews.llvm.org/D79171
Fold or(zext(bitreverse(x)),shl(zext(bitreverse(y)),bw/2) -> bitreverse(or(zext(x),shl(zext(y),bw/2))
Practically this is the same as the BSWAP pattern so we might as well handle it.
We can combine a floating-point extension cast with a conversion
from integer if we know the earlier cast is exact.
This is an optimization suggested in PR36617:
https://bugs.llvm.org/show_bug.cgi?id=36617#c19
However, this patch does not change the example suggested there.
This patch only uses the existing analysis to handle cases where
the integer source value magnitude is narrower than the
intermediate FP mantissa (guarantees that the conversion to FP is
exact). Follow-up patches to the analysis function can enable
more cases.
Differential Revision: https://reviews.llvm.org/D79116
XOP targets have fast per-element vector shifts and we're better off splitting to 128-bit shifts where necessary (which is what we already do in LowerShift).
Consider any constant memory type, not just global constants. AMDGPU
kernel parameters are effectively global constants, but appear as
either reads from an intrinsic derived pointer or function argument.
Helps improve test coverage of the XOP modes in X86TargetLowering::isVectorShiftByScalarCheap (and where we always return false for vXi8 vector shifts).
As with the extractelement patterns that are currently in vector-combine,
there are going to be several possible variations on this theme. This
should be the clearest, simplest example.
Scalarization is the right direction for target-independent canonicalization,
and InstCombine has some of those folds already, but it doesn't do this.
I proposed a similar transform in D50992. Here in vector-combine, we can
check the cost model to be sure it's profitable, so there should be less risk.
Differential Revision: https://reviews.llvm.org/D79452
Summary:
If the only use of a value is a start or end lifetime intrinsic then mark the intrinsic as trivially dead. This should allow for that value to then be removed as well.
Currently, this only works for allocas, globals, and arguments.
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79355
them in a special text section.
For sampleFDO, because the optimized build uses profile generated from
previous release, previously we couldn't tell a function without profile
was truely cold or just newly created so we had to treat them conservatively
and put them in .text section instead of .text.unlikely. The result was when
we persuing the best performance by locking .text.hot and .text in memory,
we wasted a lot of memory to keep cold functions inside.
In https://reviews.llvm.org/D66374, we introduced profile symbol list to
discriminate functions being cold versus functions being newly added.
This mechanism works quite well for regular use cases in AutoFDO. However,
in some case, we can only have a partial profile when optimizing a target.
The partial profile may be an aggregated profile collected from many targets.
The profile symbol list method used for regular sampleFDO profile is not
applicable to partial profile use case because it may be too large and
introduce many false positives.
To solve the problem for partial profile use case, we provide an option called
--profile-unknown-in-special-section. For functions without profile, we will
still treat them conservatively in compiler optimizations -- for example,
treat them as warm instead of cold in inliner. When we use profile info to
add section prefix for functions, we will discriminate functions known to be
not cold versus functions without profile (being unknown), and we will put
functions being unknown in a special text section called .text.unknown.
Runtime system will have the flexibility to decide where to put the special
section in order to achieve a balance between performance and memory saving.
Differential Revision: https://reviews.llvm.org/D62540
No changes relative to last time, but after a mitigation for
an AMDGPU regression landed.
---
If SimplifyInstruction() does not succeed in simplifying the
instruction, it will compute the known bits of the instruction
in the hope that all bits are known and the instruction can be
folded to a constant. I have removed a similar optimization
from InstCombine in D75801, and would like to drop this one as well.
On average, we spend ~1% of total compile-time performing this
known bits calculation. However, if we introduce some additional
statistics for known bits computations and how many of them succeed
in simplifying the instruction we get (on test-suite):
instsimplify.NumKnownBits: 216
instsimplify.NumKnownBitsComputed: 13828375
valuetracking.NumKnownBitsComputed: 45860806
Out of ~14M known bits calculations (accounting for approximately
one third of all known bits calculations), only 0.0015% succeed in
producing a constant. Those cases where we do succeed to compute
all known bits will get folded by other passes like InstCombine
later. On test-suite, only lencod.test and GCC-C-execute-pr44858.test
show a hash difference after this change. On lencod we see an
improvement (a loop phi is optimized away), on the GCC torture
test a regression (a function return value is determined only
after IPSCCP, preventing propagation from a noinline function.)
There are various regressions in InstSimplify tests. However, all
of these cases are already handled by InstCombine, and corresponding
tests have already been added there.
Differential Revision: https://reviews.llvm.org/D79294
This patch removes FC0.ExitBlock and FC1GuardBlock from DT and LI
after fusion of guarded loops. They become unreachable and LI
verification failed when they happened to be inside another loop.
Reviewed By: kbarton
Differential Revision: https://reviews.llvm.org/D78679
In a recent patch we introduced a problem with abstract attributes that
were assumed dead at some point. Since `Attributor::updateAA` was
introduced in 95e0d28b71, we did not
remember the dependence on the liveness AA when an abstract attribute
was assumed dead and therefore not updated.
Explicit reproducer added in liveness.ll.
---
Single run of the Attributor module and then CGSCC pass (oldPM)
for SPASS/clause.c (~10k LLVM-IR loc):
Before:
```
calls to allocation functions: 509242 (345483/s)
temporary memory allocations: 98666 (66937/s)
peak heap memory consumption: 18.60MB
peak RSS (including heaptrack overhead): 103.29MB
total memory leaked: 269.10KB
```
After:
```
calls to allocation functions: 529332 (355494/s)
temporary memory allocations: 102107 (68574/s)
peak heap memory consumption: 19.40MB
peak RSS (including heaptrack overhead): 102.79MB
total memory leaked: 269.10KB
```
Difference:
```
calls to allocation functions: 20090 (1339333/s)
temporary memory allocations: 3441 (229400/s)
peak heap memory consumption: 801.45KB
peak RSS (including heaptrack overhead): 0B
total memory leaked: 0B
```
Summary:
Update the check for the default exit block to not only check that the
terminator is not unreachable, but also check that unreachable block has
*only* the unreachable instruction.
Reviewers: chandlerc
Subscribers: hiraditya, uabelho, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78277
Summary:
This patch fix the following issues with visitExtractElementInst:
1. Restrict VectorUtils::findScalarElement to fixed-length vector.
For scalable type, the number of elements in shuffle mask is
unknown at compile-time.
2. Fix out-of-range calculation for fixed-length vector.
3. Skip scalable type when analysis rely on fixed number of elements.
4. Add unit tests to check functionality of extractelement for scalable type.
Reviewers: sdesmalen, efriedma, spatel, nikic
Reviewed By: efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78267
Summary:
This patch fixes the following issues in visitInsertElementInst:
1. Bail out for scalable type when analysis requires fixed size number of vector elements.
2. Use cast<FixedVectorType> to get vector number of elements. This ensure assertion
on scalable vector type.
3. For scalable type, avoid folding a chain of insertelement into splat:
insertelt(insertelt(insertelt(insertelt X, %k, 0), %k, 1), %k, 2) ...
->
shufflevector(insertelt(X, %k, 0), undef, zero)
The length of scalable vector is unknown at compile-time, therefore we don't know if
given insertelement sequence is valid for splat.
Reviewers: sdesmalen, efriedma, spatel, nikic
Reviewed By: sdesmalen, efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78895
The original patch (rG86dfbc676ebe) exposed an existing bug:
we could wrongly cast a constant expression to BinaryOperator
because the pattern matching allows that. This adds a check
for that case, and there's a reduced test case to verify no
crashing.
Original commit message:
This builds on the or-reduction bailout that was added with D67841.
We still do not have IR-level load combining, although that could
be a target-specific enhancement for -vector-combiner.
The heuristic is narrowly defined to catch the motivating case from
PR39538:
https://bugs.llvm.org/show_bug.cgi?id=39538
...while preserving existing functionality.
That is, there's an unmodified test of pure load/zext/store that is
not seen in this patch at llvm/test/Transforms/SLPVectorizer/X86/cast.ll.
That's the reason for the logic difference to require the 'or'
instructions. The chances that vectorization would actually help a
memory-bound sequence like that seem small, but it looks nicer with:
vpmovzxwd (%rsi), %xmm0
vmovdqu %xmm0, (%rdi)
rather than:
movzwl (%rsi), %eax
movl %eax, (%rdi)
...
In the motivating test, we avoid creating a vector mess that is
unrecoverable in the backend, and SDAG forms the expected bswap
instructions after load combining:
movzbl (%rdi), %eax
vmovd %eax, %xmm0
movzbl 1(%rdi), %eax
vmovd %eax, %xmm1
movzbl 2(%rdi), %eax
vpinsrb $4, 4(%rdi), %xmm0, %xmm0
vpinsrb $8, 8(%rdi), %xmm0, %xmm0
vpinsrb $12, 12(%rdi), %xmm0, %xmm0
vmovd %eax, %xmm2
movzbl 3(%rdi), %eax
vpinsrb $1, 5(%rdi), %xmm1, %xmm1
vpinsrb $2, 9(%rdi), %xmm1, %xmm1
vpinsrb $3, 13(%rdi), %xmm1, %xmm1
vpslld $24, %xmm0, %xmm0
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpslld $16, %xmm1, %xmm1
vpor %xmm0, %xmm1, %xmm0
vpinsrb $1, 6(%rdi), %xmm2, %xmm1
vmovd %eax, %xmm2
vpinsrb $2, 10(%rdi), %xmm1, %xmm1
vpinsrb $3, 14(%rdi), %xmm1, %xmm1
vpinsrb $1, 7(%rdi), %xmm2, %xmm2
vpinsrb $2, 11(%rdi), %xmm2, %xmm2
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpinsrb $3, 15(%rdi), %xmm2, %xmm2
vpslld $8, %xmm1, %xmm1
vpmovzxbd %xmm2, %xmm2 # xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero
vpor %xmm2, %xmm1, %xmm1
vpor %xmm1, %xmm0, %xmm0
vmovdqu %xmm0, (%rsi)
movl (%rdi), %eax
movl 4(%rdi), %ecx
movl 8(%rdi), %edx
movbel %eax, (%rsi)
movbel %ecx, 4(%rsi)
movl 12(%rdi), %ecx
movbel %edx, 8(%rsi)
movbel %ecx, 12(%rsi)
Differential Revision: https://reviews.llvm.org/D78997
If tail-folding of the scalar remainder loop is applied, the primary induction
variable is splat to a vector and used by the masked load/store vector
instructions, thus the IV does not remain scalar. Because we now mark
that the IV does not remain scalar for these cases, we don't emit the vector IV
if it is not used. Thus, the vectoriser produces less dead code.
Thanks to Ayal Zaks for the direction how to fix this.
Differential Revision: https://reviews.llvm.org/D78911
loop nest.
Summary: As discussed in https://reviews.llvm.org/D73129.
Example
Before unroll and jam:
for
A
for
B
for
C
D
E
After unroll and jam (currently):
for
A
A'
for
B
for
C
D
B'
for
C'
D'
E
E'
After unroll and jam (Ideal):
for
A
A'
for
B
B'
for
C
C'
D
D'
E
E'
This is the first patch to change unroll and jam to work in the ideal
way.
This patch change the safety checks needed to make sure is safe to
unroll and jam in the ideal way.
Reviewer: dmgreen, jdoerfert, Meinersbur, kbarton, bmahjour, etiotto
Reviewed By: Meinersbur
Subscribers: fhahn, hiraditya, zzheng, llvm-commits, anhtuyen, prithayan
Tag: LLVM
Differential Revision: https://reviews.llvm.org/D76132
Summary:
If the only use of a value is a start or end lifetime intrinsic then mark the intrinsic as trivially dead. This should allow for that value to then be removed as well.
Currently, this only works for allocas, globals, and arguments.
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79355
Summary:
getLogBase2 tries to iterate over the number of vector elements. Since
the number of elements of a scalable vector is unknown at compile time,
we must return null if the input type is scalable.
Identified by test LLVM.Transforms/InstCombine::nsw.ll
Reviewers: efriedma, fpetrogalli, kmclaughlin, spatel
Reviewed By: efriedma, fpetrogalli
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D79197
The 'nsz' flag is different than 'nnan' or 'ninf' in that it does not create poison.
Make that explicit in the LangRef and fix ValueTracking analysis that misinterpreted
the definition.
This manifests as bugs in InstSimplify shown in the test diffs and as discussed in
PR45778:
https://bugs.llvm.org/show_bug.cgi?id=45778
Differential Revision: https://reviews.llvm.org/D79422
This builds on the or-reduction bailout that was added with D67841.
We still do not have IR-level load combining, although that could
be a target-specific enhancement for -vector-combiner.
The heuristic is narrowly defined to catch the motivating case from
PR39538:
https://bugs.llvm.org/show_bug.cgi?id=39538
...while preserving existing functionality.
That is, there's an unmodified test of pure load/zext/store that is
not seen in this patch at llvm/test/Transforms/SLPVectorizer/X86/cast.ll.
That's the reason for the logic difference to require the 'or'
instructions. The chances that vectorization would actually help a
memory-bound sequence like that seem small, but it looks nicer with:
vpmovzxwd (%rsi), %xmm0
vmovdqu %xmm0, (%rdi)
rather than:
movzwl (%rsi), %eax
movl %eax, (%rdi)
...
In the motivating test, we avoid creating a vector mess that is
unrecoverable in the backend, and SDAG forms the expected bswap
instructions after load combining:
movzbl (%rdi), %eax
vmovd %eax, %xmm0
movzbl 1(%rdi), %eax
vmovd %eax, %xmm1
movzbl 2(%rdi), %eax
vpinsrb $4, 4(%rdi), %xmm0, %xmm0
vpinsrb $8, 8(%rdi), %xmm0, %xmm0
vpinsrb $12, 12(%rdi), %xmm0, %xmm0
vmovd %eax, %xmm2
movzbl 3(%rdi), %eax
vpinsrb $1, 5(%rdi), %xmm1, %xmm1
vpinsrb $2, 9(%rdi), %xmm1, %xmm1
vpinsrb $3, 13(%rdi), %xmm1, %xmm1
vpslld $24, %xmm0, %xmm0
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpslld $16, %xmm1, %xmm1
vpor %xmm0, %xmm1, %xmm0
vpinsrb $1, 6(%rdi), %xmm2, %xmm1
vmovd %eax, %xmm2
vpinsrb $2, 10(%rdi), %xmm1, %xmm1
vpinsrb $3, 14(%rdi), %xmm1, %xmm1
vpinsrb $1, 7(%rdi), %xmm2, %xmm2
vpinsrb $2, 11(%rdi), %xmm2, %xmm2
vpmovzxbd %xmm1, %xmm1 # xmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero
vpinsrb $3, 15(%rdi), %xmm2, %xmm2
vpslld $8, %xmm1, %xmm1
vpmovzxbd %xmm2, %xmm2 # xmm2 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero
vpor %xmm2, %xmm1, %xmm1
vpor %xmm1, %xmm0, %xmm0
vmovdqu %xmm0, (%rsi)
movl (%rdi), %eax
movl 4(%rdi), %ecx
movl 8(%rdi), %edx
movbel %eax, (%rsi)
movbel %ecx, 4(%rsi)
movl 12(%rdi), %ecx
movbel %edx, 8(%rsi)
movbel %ecx, 12(%rsi)
Differential Revision: https://reviews.llvm.org/D78997
We check that C is finite and strictly positive, but there's no need to
check that it's normal too. exp2 should be just as accurate on denormals
as pow is.
Differential Revision: https://reviews.llvm.org/D79413
I don't think there's any good reason not to do this transformation when
the pow has multiple uses.
Differential Revision: https://reviews.llvm.org/D79407
This adds a general combine that can be used to fold:
or(zext(OP(x)), shl(zext(OP(y)),bw/2))
-->
OP(or(zext(x), shl(zext(y),bw/2)))
Allowing us to widen 'concat-able' style or+zext patterns - I've just set this up for BSWAP but we could use this for other similar ops (BITREVERSE for instance).
We already do something similar for bitop(bswap(x),bswap(y)) --> bswap(bitop(x,y))
Fixes PR45715
Reviewed By: @lebedev.ri
Differential Revision: https://reviews.llvm.org/D79041
optimizePow does not create any new calls to pow, so it should work
regardless of whether the pow library function is available. This allows
it to optimize the llvm.pow intrinsic on targets with no math library.
Based on a patch by Tim Renouf.
Differential Revision: https://reviews.llvm.org/D68231
Before we eagerly put dependences into the QueryMap as soon as we
encountered them (via `Attributor::getAAFor<>` or
`Attributor::recordDependence`). Now we will wait to see if the
dependence is useful, that is if the target is not already in a fixpoint
state at the end of the update. If so, there is no need to record the
dependence at all.
Due to the abstraction via `Attributor::updateAA` we will now also treat
the very first update (during attribute creation) as we do subsequent
updates.
Finally this resolves the problematic usage of QueriedNonFixAA.
---
Single run of the Attributor module and then CGSCC pass (oldPM)
for SPASS/clause.c (~10k LLVM-IR loc):
Before:
```
calls to allocation functions: 554675 (389245/s)
temporary memory allocations: 101574 (71280/s)
peak heap memory consumption: 28.46MB
peak RSS (including heaptrack overhead): 116.26MB
total memory leaked: 269.10KB
```
After:
```
calls to allocation functions: 512465 (345559/s)
temporary memory allocations: 98832 (66643/s)
peak heap memory consumption: 22.54MB
peak RSS (including heaptrack overhead): 106.58MB
total memory leaked: 269.10KB
```
Difference:
```
calls to allocation functions: -42210 (-727758/s)
temporary memory allocations: -2742 (-47275/s)
peak heap memory consumption: -5.92MB
peak RSS (including heaptrack overhead): 0B
total memory leaked: 0B
```
Attributes that only depend on the value (=bit pattern) can be
initialized from uses in the must-be-executed-context (MBEC). We did use
`AAComposeTwoGenericDeduction` and `AAFromMustBeExecutedContext` before
to do this for some positions of these attributes but not for all. This
was fairly complicated and also problematic as we did run it in every
`updateImpl` call even though we only use known information. The new
implementation removes `AAComposeTwoGenericDeduction`* and
`AAFromMustBeExecutedContext` in favor of a simple interface
`AddInformation::fromMBEContext(...)` which we call from the
`initialize` methods of the "value attribute" `Impl` classes, e.g.
`AANonNullImpl:initialize`.
There can be two types of test changes:
1) Artifacts were we miss some information that was known before a
global fixpoint was reached and therefore available in an update
but not at the beginning.
2) Deduction for values we did not derive via the MBEC before or which
were not found as the `AAFromMustBeExecutedContext::updateImpl` was
never invoked.
* An improved version of AAComposeTwoGenericDeduction can be found in
D78718. Once we find a new use case that implementation will be able
to handle "generic" AAs better.
---
Single run of the Attributor module and then CGSCC pass (oldPM)
for SPASS/clause.c (~10k LLVM-IR loc):
Before:
```
calls to allocation functions: 468428 (328952/s)
temporary memory allocations: 77480 (54410/s)
peak heap memory consumption: 32.71MB
peak RSS (including heaptrack overhead): 122.46MB
total memory leaked: 269.10KB
```
After:
```
calls to allocation functions: 554720 (351310/s)
temporary memory allocations: 101650 (64376/s)
peak heap memory consumption: 28.46MB
peak RSS (including heaptrack overhead): 116.75MB
total memory leaked: 269.10KB
```
Difference:
```
calls to allocation functions: 86292 (556722/s)
temporary memory allocations: 24170 (155935/s)
peak heap memory consumption: -4.25MB
peak RSS (including heaptrack overhead): 0B
total memory leaked: 0B
```
Reviewed By: uenoku
Differential Revision: https://reviews.llvm.org/D78719
If SimplifyInstruction() does not succeed in simplifying the
instruction, it will compute the known bits of the instruction
in the hope that all bits are known and the instruction can be
folded to a constant. I have removed a similar optimization
from InstCombine in D75801, and would like to drop this one as well.
On average, we spend ~1% of total compile-time performing this
known bits calculation. However, if we introduce some additional
statistics for known bits computations and how many of them succeed
in simplifying the instruction we get (on test-suite):
instsimplify.NumKnownBits: 216
instsimplify.NumKnownBitsComputed: 13828375
valuetracking.NumKnownBitsComputed: 45860806
Out of ~14M known bits calculations (accounting for approximately
one third of all known bits calculations), only 0.0015% succeed in
producing a constant. Those cases where we do succeed to compute
all known bits will get folded by other passes like InstCombine
later. On test-suite, only lencod.test and GCC-C-execute-pr44858.test
show a hash difference after this change. On lencod we see an
improvement (a loop phi is optimized away), on the GCC torture
test a regression (a function return value is determined only
after IPSCCP, preventing propagation from a noinline function.)
There are various regressions in InstSimplify tests. However, all
of these cases are already handled by InstCombine, and corresponding
tests have already been added there.
Differential Revision: https://reviews.llvm.org/D79294
When the shufflevector mask operand was converted into special
instruction data, the FunctionComparator was not updated to
account for this. As such, MergeFuncs will happily merge
shufflevectors with different masks.
This fixes https://bugs.llvm.org/show_bug.cgi?id=45773.
Differential Revision: https://reviews.llvm.org/D79261
The more general fold was not poison-safe, so it was removed:
rG5486e00
...but it is ok to have this transform if analysis can determine
the vector contains no poison. The test shows a simple example
of that: constant integer elements are not poison.
PR45481:
https://bugs.llvm.org/show_bug.cgi?id=45481
SDAG has an identical transform to this, so there's little
chance of any real-world impact. OTOH, that means we are
effectively sweeping the bug out of sight because poison
exists in codegen too.
In D74183 clang started emitting alignment for sret parameters
unconditionally. This caused a 1.5% compile-time regression on
tramp3d-v4. The reason is that we now generate many instance of IR like
%ptrint = ptrtoint %class.GuardLayers* %guards_m to i64
%maskedptr = and i64 %ptrint, 3
%maskcond = icmp eq i64 %maskedptr, 0
tail call void @llvm.assume(i1 %maskcond)
to preserve the alignment information during inlining. Based on IR
analysis, these assumptions also regress optimization. The attached
phase ordering test case illustrates two issues: One are instruction
count based optimization heuristics, which are affected by the four
additional instructions of the assumption. The other is blocking of
SROA due to ptrtoint casts (PR45763).
We already encountered the same problem in Rust, where we (unlike
Clang) generally prefer to emit alignment information absolutely
everywhere it is available. We were only able to do this after
hardcoding -preserve-alignment-assumptions-during-inlining=false,
because we were seeing significant optimization and compile-time
regressions otherwise.
This patch disables -preserve-alignment-assumptions-during-inlining
by default, because we should not be punishing people for adding
more alignment annotations.
Once the assume bundle work shakes out and we can represent (and use)
alignment assumptions using assume bundles, it should be possible to
re-enable this with reduced overhead.
Differential Revision: https://reviews.llvm.org/D76886
The assert checks that every instruction must be annotated by this point while it is not
necessary. If the inlining process was interrupted because the threshold was reached, the rest
of the instructions would not be annotated which triggers the assert.
The added test shows the situation in which it can happen.
This is a recommit as the original commit fail due to the absence of REQUIRES: assert in the test.
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D79107
Summary: Currenlty BPI unconditionally creates post dominator tree each time. While this is not incorrect we can save compile time by reusing existing post dominator tree (when it's valid) provided by analysis manager.
Reviewers: skatkov, taewookoh, yrouban
Reviewed By: skatkov
Subscribers: hiraditya, steven_wu, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78987
The assert checks that every instruction must be annotated by this point while it is not
necessary. If the inlining process was interrupted because the threshold was reached, the rest
of the instructions would not be annotated which triggers the assert.
The added test shows the situation in which it can happen.
Reviewed-By: mtrofin
Diff: https://reviews.llvm.org/D79107
The improvements to the x86 vector insert/extract element costs in D74976 resulted in the estimated costs for vector initialization and scalarization increasing higher than should be expected. This is particularly noticeable on pre-SSE4 targets where the available of legal INSERT_VECTOR_ELT ops is more limited.
This patch does 2 things:
1 - it implements X86TTIImpl::getScalarizationOverhead to more accurately represent the typical costs of a ISD::BUILD_VECTOR pattern.
2 - it adds a DemandedElts mask to getScalarizationOverhead to permit the SLP's BoUpSLP::getGatherCost to be rewritten to use it directly instead of accumulating raw vector insertion costs.
This fixes PR45418 where a v4i8 (zext'd to v4i32) was no longer vectorizing.
A future patch should extend X86TTIImpl::getScalarizationOverhead to tweak the EXTRACT_VECTOR_ELT scalarization costs as well.
Reviewed By: @craig.topper
Differential Revision: https://reviews.llvm.org/D78216
The crash that caused the original revert has been fixed in
a3c964a278. I also added a reduced version of the crash reproducer.
This reverts the revert commit 2107af9ccf.
In `InstCombiner::visitAdd()`, we have
```
// A+B --> A|B iff A and B have no bits set in common.
if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
return BinaryOperator::CreateOr(LHS, RHS);
```
so we should handle such `or`'s here, too.
Profile and profile summary are usually read only once and then annotated
on IR. The profile summary metadata on IR should include the value of the
newly added partial profile flag, so that compilation phase like thinlto
postlink can get the full set of profile information.
Differential Revision: https://reviews.llvm.org/D78310
When folding tail, branch taken count is computed during initial VPlan execution
and recorded to be used by the compare computing the loop's mask. This recording
should directly set the State, instead of reusing Value2VPValue mapping which
serves original Values present prior to vectorization.
The branch taken count may be a constant Value, which may be used elsewhere in
the loop; trying to employ Value2VPValue for both leads to the issue reported in
https://reviews.llvm.org/D76992#inline-721028
Differential Revision: https://reviews.llvm.org/D78847
Integer ranges can be used for loaded/stored values. Note that widening
can be disabled for loads/stores, as we only rely on instructions that
cause continued increases to ranges to be widened (like binary
operators).
Reviewers: efriedma, mssimpso, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D78433
(X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
(X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
We have more analyis for 'and' patterns and already lean this way
in the existing code, so this should be neutral or better in IR.
If this does not do as well in codegen, the problem already exists
and we should fix that based on target costs/heuristics.
http://volta.cs.utah.edu:8080/z/oP3ecL
define void @src(i8 %x, i8 %OrC, i8 %C, i1* %p0, i1* %p1) {
%or = or i8 %x, %OrC
%eq = icmp eq i8 %or, %C
store i1 %eq, i1* %p0
%ne = icmp ne i8 %or, %C
store i1 %ne, i1* %p1
ret void
}
define void @tgt(i8 %x, i8 %OrC, i8 %C, i1* %p0, i1* %p1) {
%NotOrC = xor i8 %OrC, -1
%a = and i8 %x, %NotOrC
%NewC = xor i8 %C, %OrC
%eq = icmp eq i8 %a, %NewC
store i1 %eq, i1* %p0
%ne = icmp ne i8 %a, %NewC
store i1 %ne, i1* %p1
ret void
}
Using the existing NumFastStores statistic can be misleading when
comparing the impact of DSE patches.
For example, consider the case where a store gets removed from a
function before it is inlined into another function. A less
powerful DSE might only remove the store from functions it has
been inlined into, which will result in more stores being removed, but
no difference in the actual number of stores after DSE.
The new stat provides the absolute number of stores surviving after
DSE.
Reviewers: dmgreen, bryant, asbirlea, jfb
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D78830
Summary:
This patch helps isGuaranteedNotToBeUndefOrPoison look into more constants and instructions (bitcast/alloca/gep/fcmp).
To deal with bitcast, Depth is added to isGuaranteedNotToBeUndefOrPoison.
This patch is splitted from https://reviews.llvm.org/D75808.
Checked with Alive2
Reviewers: reames, jdoerfert
Reviewed By: jdoerfert
Subscribers: sanwou01, spatel, llvm-commits, hiraditya
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76010
Currently an unknown/undef value is marked as overdefined when merged
with an empty range. An empty range can occur in unreachable/dead code.
When merging the new unknown state (= no value known yet) with an empty
range, there still isn't any information about the value yet and we can
stay in unknown.
This gives a few nice improvements on the number of instructions removed
by IPSCCP:
Same hash: 170 (filtered out)
Remaining: 67
Metric: sccp.IPNumInstRemoved
Program base patch diff
test-suite...rks/FreeBench/mason/mason.test 3.00 6.00 100.0%
test-suite...nchmarks/McCat/18-imp/imp.test 3.00 5.00 66.7%
test-suite...C/CFP2000/179.art/179.art.test 2.00 3.00 50.0%
test-suite...ijndael/security-rijndael.test 2.00 3.00 50.0%
test-suite...ks/Prolangs-C/agrep/agrep.test 40.00 58.00 45.0%
test-suite...ce/Applications/Burg/burg.test 26.00 37.00 42.3%
test-suite...cCat/03-testtrie/testtrie.test 3.00 4.00 33.3%
test-suite...Source/Benchmarks/sim/sim.test 29.00 36.00 24.1%
test-suite.../Applications/spiff/spiff.test 9.00 11.00 22.2%
test-suite...s/FreeBench/neural/neural.test 5.00 6.00 20.0%
test-suite...pplications/treecc/treecc.test 66.00 79.00 19.7%
test-suite...langs-C/football/football.test 85.00 101.00 18.8%
test-suite...ce/Benchmarks/PAQ8p/paq8p.test 90.00 105.00 16.7%
test-suite...oxyApps-C++/miniFE/miniFE.test 37.00 43.00 16.2%
test-suite...rks/FreeBench/pifft/pifft.test 26.00 30.00 15.4%
test-suite...lications/sqlite3/sqlite3.test 481.00 548.00 13.9%
test-suite...marks/7zip/7zip-benchmark.test 4875.00 5522.00 13.3%
test-suite.../CINT2000/176.gcc/176.gcc.test 1117.00 1197.00 7.2%
test-suite...0.perlbench/400.perlbench.test 1618.00 1732.00 7.0%
Reviewers: efriedma, nikic, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D78667
Summary:
refactor assume bulider for the next patch.
the assume builder now generate only one assume per attribute kind and per value they are on. to do this it takes the highest. this is desirable because currently, for all attributes the higest value is the most valuable.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78013
We should only skip `lifetime` and `dbg` intrinsics when searching for users.
Other intrinsics are legit users that can't be ignored.
Without this fix, the testcase would result in an invalid IR. `memcpy`
will have a reference to the, now, external value (local to the
extracted loop function).
Fix PR42194
Differential Revision: https://reviews.llvm.org/D78749
If one of caller/callee has disabled ZMM registers due to
prefer-vector-width=256, we were previously
disabling argument promotion as the ABI might be incompatible since
one side will split 512-bit vectors in this case.
But if we can see that the types are all scalar this shouldn't be
a problem.
This patch assumes that pointer element type reflects the type that
the argument will be promoted to.
Differential Revision: https://reviews.llvm.org/D78770
Summary:
This simplifies testing in scenarios where we want to set up module-wide
analyses for inlining. The patch enables treating inlining and its
function cleanups, as a module pass. The alternative would be for tests
to describe the pipeline, which is tedious and adds maintenance
overhead.
Reviewers: davidxl, dblaikie, jdoerfert, sstefan1
Subscribers: hiraditya, steven_wu, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78512
As discussed in PR45478:
https://bugs.llvm.org/show_bug.cgi?id=45478
...propagating FMF from the outer (second) call is not correct,
so intersect them instead.
I suspect we could do better (see TODO comment), but mismatched
FMF is probably too rare to care about.
Differential Revision: https://reviews.llvm.org/D78631
One of transforms the loop vectorizer makes is LCSSA formation. In some cases it
is the only transform it makes. We should not drop CFG analyzes if only LCSSA was
formed and no actual CFG changes was made.
We should think of expanding this logic to other passes as well, and maybe make
it a part of PM framework.
Reviewed By: Florian Hahn
Differential Revision: https://reviews.llvm.org/D78360
The three tests modified by this commit have been partially disabled
(one run line is commented out). As a consequence subsequent updates
will have weird effects on the check lines. This is a commit to avoid
such effects by making the check lines match the three remaining run
lines.
While we can do that, it doesn't increase instruction count,
if the old `sub` sticks around then the transform is not only
not a unlikely win, but a likely regression, since we likely
now extended live range and use count of both of the `sub` operands,
as opposed to just the result of `sub`.
As Kostya Serebryany notes in post-commit review in
https://reviews.llvm.org/D68408#1998112
this indeed can degrade final assembly,
increase register pressure, and spilling.
This isn't what we want here,
so at least for now let's guard it with an use check.
This was backwards from intended and missing a test. We perhaps should
just ignored the FP mode here, since it shouldn't be legal to mix code
with different default modes in the absence of strictfp.
visitExtractValueInst uses mergeInValue, so it already can handle
constant ranges. Initially the early exit was using isOverdefined to
keep things as NFC during the initial move to ValueLatticeElement.
As the function already supports constant ranges, it can just use
ValueState[&I].isOverdefined.
Reviewers: efriedma, mssimpso, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D78393
Summary:
When an irreducible SCC is converted into a new natural loop, existing
loops included in that SCC now become children of the new loop. The
logic that moves these loops from the parent loop to the new loop
invoked undefined behaviour when it modified the container that it was
iterating over. Fixed this by first extracting all the loops that are
to be removed from the parent.
Fixes bug 45623.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D78544
If we have a dependence between an abstract attribute A to an abstract
attribute B such hat changes in A should trigger an update of B, we do
not need to keep the dependence around once the update was triggered. If
the dependence is still required the update will reinsert it into the
dependence map, if it is not we avoid triggering B in the future. This
replaces the "recompute interval" mechanism we used before to prune
stale dependences.
Number of required iterations is generally down, compile time for the
module pass (not really the CGSCC pass) is down quite a bit.
There is one test change which looks like an artifact in the undefined
behavior AA that needs to be looked at.
Summary:
As we have discussed previously (e.g. in D63992 / D64090 / [[ https://bugs.llvm.org/show_bug.cgi?id=42457 | PR42457 ]]), `sub` instruction
can almost be considered non-canonical. While we do convert `sub %x, C` -> `add %x, -C`,
we sparsely do that for non-constants. But we should.
Here, i propose to interpret `sub %x, %y` as `add (sub 0, %y), %x` IFF the negation can be sinked into the `%y`
This has some potential to cause endless combine loops (either around PHI's, or if there are some opposite transforms).
For former there's `-instcombine-negator-max-depth` option to mitigate it, should this expose any such issues
For latter, if there are still any such opposing folds, we'd need to remove the colliding fold.
In any case, reproducers welcomed!
Reviewers: spatel, nikic, efriedma, xbolva00
Reviewed By: spatel
Subscribers: xbolva00, mgorny, hiraditya, reames, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68408
These are mostly replicated from D78430 (instsimplify).
If we implement more general transforms for instcombine,
then we probably don't need to add that complexity to instsimplify.
Before we kept the first applicable `ident_t*` during deduplication of
runtime calls. The problem is that "first" is dependent on the iteration
order of a DenseMap. Since the proper solution, which is to combine the
information from all `ident_t*`, should be deterministic on its own, we
will not try to make the iteration order deterministic. Instead, we will
create a fresh `ident_t*` if there is not a unique existing `ident_t*`
to pick.
With clang option -funique-internal-linkage-symbols, symbols with
internal linkage get names with the module hash appended.
Differential Revision: https://reviews.llvm.org/D78243
This will allow us to use the datalayout to disambiguate other
constructs in IR, like load alignment. Split off from D78403.
Differential Revision: https://reviews.llvm.org/D78413
Summary:
The indexing operator in Scatterer may result in building new
instructions. When using multiple such operators in a function
argument list the order in which we build instructions depend on
argument evaluation order (which is undefined in C++).
This patch avoid such problems by expanding the components using
the [] operator prior to the function call.
Problem was seen when comparing output, while builing LLVM with
different compilers (clang vs gcc).
Reviewers: foad, cameron.mcinally, uabelho
Reviewed By: foad
Subscribers: hiraditya, mgrang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78455
bitcast (shuf V, MaskC) --> shuf (bitcast V), MaskC'
This is the widen shuffle elements enhancement to D76727.
It builds on the analysis and simplifications in
D77881 and rG6a7e958a423e.
The phase ordering tests show that we can simplify inverse
shuffles across a binop in both directions (widen/narrow or
narrow/widen) now.
There's another potential transform visible in some of the
remaining TODOs - move a bitcasted operand of a shuffle
after the shuffle.
Differential Revision: https://reviews.llvm.org/D78371
See PR45510:
https://bugs.llvm.org/show_bug.cgi?id=45510
We had partial coverage for some of these patterns, so removing duplicate tests
with the complete set in the new test file.
First-order recurrences require special treatment when they are live-out;
such treatment is provided by fixFirstOrderRecurrence(), so they should be
included in AllowedExit set.
(Should probably have been included originally in D16197.)
Fixes PR45526: AllowedExit set is used by prepareToFoldTailByMasking() to
check whether the treatment for live-outs also holds when folding the tail,
which is not (yet) the case for first-order recurrences.
Differential Revision: https://reviews.llvm.org/D78210
Cost-modeling decisions are tied to the compute interleave groups
(widening decisions, scalar and uniform values). When invalidating the
interleave groups, those decisions also need to be invalidated.
Otherwise there is a mis-match during VPlan construction.
VPWidenMemoryRecipes created initially are left around w/o converting them
into VPInterleave recipes. Such a conversion indeed should not take place,
and these gather/scatter recipes may in fact be right. The crux is leaving around
obsolete CM_Interleave (and dependent) markings of instructions along with
their costs, instead of recalculating decisions, costs, and recipes.
Alternatively to forcing a complete recompute later on, we could try
to selectively invalidate the decisions connected to the interleave
groups. But we would likely need to run the uniform/scalar value
detection parts again anyways and the extra complexity is probably not
worth it.
Fixes PR45572.
Reviewers: gilr, rengolin, Ayal, hsaito
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D78298
I've always found the "findValue" a little odd and
inconsistent with other things in SDB.
This simplfifies the code in SDB to just handle a splat constant
address or a 2 operand GEP in the same BB. This removes the
need for "findValue" since the operands to the GEP are
guaranteed to be available. The splat constant handling is
new, but was needed to avoid regressions due to constant
folding combining GEPs created in CGP.
CGP is now responsible for canonicalizing gather/scatters into
this form. The pattern I'm using for scalarizing, a scalar GEP
followed by a GEP with an all zeroes index, seems to be subject
to constant folding that the insertelement+shufflevector was not.
Differential Revision: https://reviews.llvm.org/D76947
Summary:
Changes the type of the @__typeid_.*_unique_member imports we generate
for unique return value optimization from i8 to [0 x i8]. This
prevents assuming that these imports do not alias, such as when
two unique return values occur in the same vtable.
Fixes PR45393.
Reviewers: tejohnson, pcc
Reviewed By: pcc
Subscribers: aganea, hiraditya, rnk, george.burgess.iv, dblaikie, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77421
Since we use the fact that some uses are droppable in the Attributor we
need to handle them explicitly when we replace uses. As an example, an
assumed dead value can have live droppable users. In those we cannot
replace the value simply by an undef. Instead, we either drop the uses
(via `dropDroppableUses`) or keep them as they are. In this patch we do
both, depending on the situation. For values that are dead but not
necessarily removed we keep droppable uses around because they contain
information we might be able to use later. For values that are removed
we drop droppable uses explicitly to avoid replacement with undef.
The handling of the `returned` attribute in D75815 did miss the case
where the argument is (bit)casted to a different type. This is
explicitly allowed by the language reference and exposed by the
Attributor.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D77977
The check if globals were accessed was not always working because two
bits are set for NO_GLOBAL_MEM. The new check works also if only on kind
of globals (internal/external) is accessed.
When the Attributor was created the test update scripts were not well
suited to deal with the challenges of IR attribute checking. This
partially improved.
Since then we also added three additional configurations that need
testing; in total we now have the following four:
{ TUNIT, CGSCC } x { old pass manager (OPM), new pass manager (NPM) }
Finally, the number of developers and tests grew rapidly (partially due
to the addition of ArgumentPromotion and IPConstantProp tests), which
resulted in tests only being run in some configurations, different
prefixes being used, and different "styles" of checks being used.
Due to the above reasons I believed we needed to take another look at
the test update scripts. While we started to use them, via UTC_ARGS:
--enable/disable, the other problems remained. To improve the testing
situation for *all* configurations, to simplify future updates to the
test, and to help identify subtle effects of future changes, we now use
the test update scripts for (almost) all Attributor tests.
An exhaustive prefix list minimizes the number of check lines and makes
it easy to identify and compare configurations.
Tests have been adjusted in the process but we tried to keep their
intend unchanged.
Reviewed By: sstefan1
Differential Revision: https://reviews.llvm.org/D76588
The current strategy LICM uses when sinking for debuginfo is
that of picking the debug location of one of the uses.
This causes stepping to be wrong sometimes, see, e.g. PR45523.
This patch introduces a generalization of getMergedLocation(),
that operates on a vector of locations instead of two, and try
to merge all them together, and use the new API in LICM.
<rdar://problem/61750950>
We can eliminate MemoryDefs of objects not accessible after the function
returns (e.g. alloca), if there are no reads between the MemoryDef and
any function exits. We can stop traversing paths that completely
overwrite the memory location of the MemoryDef.
This patch was split off D73763.
Reviewers: dmgreen, bryant, asbirlea, Tyker, efriedma, george.burgess.iv
Reviewed By: asbirlea, george.burgess.iv
Differential Revision: https://reviews.llvm.org/D77736
An irreducible SCC is one which has multiple "header" blocks, i.e., blocks
with control-flow edges incident from outside the SCC. This pass converts an
irreducible SCC into a natural loop by introducing a single new header
block and redirecting all the edges on the original headers to this
new block.
This is a useful workaround for a limitation in the structurizer
which, which produces incorrect control flow in the presence of
irreducible regions. The AMDGPU backend provides an option to
enable this pass before the structurizer, which may eventually be
enabled by default.
Reviewed By: nhaehnle
Differential Revision: https://reviews.llvm.org/D77198
This restores commit 2ada8e2525.
Originally reverted with commit 44e09b59b8.
Fix an assert introduced in 41ed5d856c1: a phi with a single predecessor and a
mask is a valid case which is already supported by the code.
Differential Revision: https://reviews.llvm.org/D78115
This reverts commit 2ada8e2525.
Buildbots produced compilation errors which I was not able to quickly
reproduce locally. Need more time to investigate.
An irreducible SCC is one which has multiple "header" blocks, i.e., blocks
with control-flow edges incident from outside the SCC. This pass converts an
irreducible SCC into a natural loop by introducing a single new header
block and redirecting all the edges on the original headers to this
new block.
This is a useful workaround for a limitation in the structurizer
which, which produces incorrect control flow in the presence of
irreducible regions. The AMDGPU backend provides an option to
enable this pass before the structurizer, which may eventually be
enabled by default.
Reviewed By: nhaehnle
Differential Revision: https://reviews.llvm.org/D77198
Summary:
Currently, the internal options -vectorize-loops, -vectorize-slp, and
-interleave-loops do not have much practical effect. This is because
they are used to initialize the corresponding flags in the pass
managers, and those flags are then unconditionally overwritten when
compiling via clang or via LTO from the linkers. The only exception was
-vectorize-loops via opt because of some special hackery there.
While vectorization could still be disabled when compiling via clang,
using -fno-[slp-]vectorize, this meant that there was no way to disable
it when compiling in LTO mode via the linkers. This only affected
ThinLTO, since for regular LTO vectorization is done during the compile
step for scalability reasons. For ThinLTO it is invoked in the LTO
backends. See also the discussion on PR45434.
This patch makes it so the internal options can actually be used to
disable these optimizations. Ultimately, the best long term solution is
to mark the loops with metadata (similar to the approach used to fix
-fno-unroll-loops in D77058), but this enables a shorter term
workaround, and actually makes these internal options useful.
I constant propagated the initial values of these internal flags into
the pass manager flags (for some reasons vectorize-loops and
interleave-loops were initialized to true, while vectorize-slp was
initialized to false). As mentioned above, they are overwritten
unconditionally so this doesn't have any real impact, and these initial
values aren't particularly meaningful.
I then changed the passes to check the internl values and return without
performing the associated optimization when false (I changed the default
of -vectorize-slp to true so the options behave similarly). I was able
to remove the hackery in opt used to get -vectorize-loops=false to work,
as well as a special option there used to disable SLP vectorization.
Finally, I changed thinlto-slp-vectorize-pm.c to:
a) Only test SLP (moved the loop vectorization checking to a new test).
b) Use code that is slp vectorized when it is enabled, and check that
instead of whether the pass is enabled.
c) Test the new behavior of -vectorize-slp.
d) Test both pass managers.
The loop vectorization (and associated interleaving) testing I moved to
a new thinlto-loop-vectorize-pm.c test, with several changes:
a) Changed the flags on the interleaving testing so that it will
actually interleave, and check that.
b) Test the new behavior of -vectorize-loops and -interleave-loops.
c) Test both pass managers.
Reviewers: fhahn, wmi
Subscribers: hiraditya, steven_wu, dexonsmith, cfe-commits, davezarzycki, llvm-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D77989
Summary:
This patch fix the following issues in InstCombiner::visitGetElementPtrInst
1. Skip for scalable type if transformation requires fixed size number of
vector element.
2. Skip for scalable type if transformation relies on compile-time known type
alloc size.
3. Use VectorType::getElementCount when scalable property is used to construct
new VectorType.
4. Use TypeSize::getKnownMinSize when minimal size of a scalable type is valid to determine GEP 'inbounds'.
5. Explicitly call TypeSize::getFixedSize to avoid implicit type conversion to uint64_t.
Reviewers: sdesmalen, efriedma, spatel, ctetreau
Reviewed By: efriedma
Subscribers: tschuett, hiraditya, rkruppe, psnobl, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78081
Summary:
AbstractCallSite::getCallbackUses() does not check that callback callee index from
the callback metadata does not exceed the total number of call arguments. This patch
add such validation check.
Reviewers: jdoerfert, sstefan1
Reviewed By: jdoerfert
Subscribers: hiraditya, arphaman, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78112
D77635 added support to recognise primary induction variables for counting-down
loops. This allows us to fold the scalar tail loop into the main vector body,
which we need for MVE tail-predication. This adds some ARM tail-folding test
cases that we want to support.
This test was extracted from D76838, which implemented a different approach to
reverse and thus find a primary induction variable.
This patch fixes 2 related bugs in ADCE:
- `performDeadCodeElimination` does not report changes if it did ONLY
CFG changes (affects both old and new pass managers);
- When control flow removal is enabled, new pass manager does not
drop CFG analyses.
Both can lead to incorrect loop info after ADCE that does only CFG changes.
Differential Revision: https://reviews.llvm.org/D78103
Reviewed By: Denis Antrushin
This is equivalent in terms of LLVM IR semantics, but we want to
transition away from using MaybeAlign to represent the alignment of
these instructions.
Differential Revision: https://reviews.llvm.org/D77984
This reverts commit 60c642e74b.
This patch is making the TLI "closed" for a predefined set of VecLib
while at the moment it is extensible for anyone to customize when using
LLVM as a library.
Reverting while we figure out a way to re-land it without losing the
generality of the current API.
Differential Revision: https://reviews.llvm.org/D77925
Revision a1c05fe <https://reviews.llvm.org/rGa1c05fe20f3def1f1be9f50d2adefc6b6f1578ad>
removed bitcast from the list of problematic transformations, however:
%97 = fptrunc ppc_fp128 %2 to double // we need to check ppc_fp128 here to prevent the transformation
%98 = bitcast double %97 to i64 // a1c05fe checks ppc_fp128 at here
%99 = icmp slt i64 %98, 0
%100 = zext i1 %99 to i8
store i8 %100, i8* %7, align 1
so this patch does that. I'm also disabling it in the presence of extend just in case.
I verified separately that the hash of -std::infinity and std::infinity don't match now.
Differential Revision: https://reviews.llvm.org/D77911
For non-integer constants/expressions and overdefined, I think we can
just use SimplifyBinOp to do common folds. By just passing a context
with the DL, SimplifyBinOp should not try to get additional information
from looking at definitions.
For overdefined values, it should be enough to just pass the original
operand.
Note: The comment before the `if (isconstant(V1State)...` was wrong
originally: isConstant() also matches integer ranges with a single
element. It is correct now.
Reviewers: efriedma, davide, mssimpso, aartbik
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D76459
Loop simplify form should always be checked because logic of
propagateStoredValueToLoadUsers relies on it (in particular, it
requires preheader).
Reviewed By: Fedor Sergeev, Florian Hahn
Differential Revision: https://reviews.llvm.org/D77775
Summary:
Encode `-fveclib` setting as per-function attribute so it can threaded through to LTO backends. Accordingly per-function TLI now reads
the attributes and select available vector function list based on that. Now we also populate function list for all supported vector
libraries for the shared per-module `TargetLibraryInfoImpl`, so each function can select its available vector list independently but without
duplicating the vector function lists. Inlining between incompatbile vectlib attributed is also prohibited now.
Subscribers: hiraditya, dexonsmith, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D77632
This is similar to what I recently did for getArithmeticReductionCost.
I'm trying to account for the narrowing from 512->256->128 as we go.
I've also added a new helper method getMinMaxCost that tries to
handle the cases where we have native min/max instructions and
fall back to cmp+select when we don't.
Differential Revision: https://reviews.llvm.org/D76634
Introduce a new VPWidenCanonicalIVRecipe to generate a canonical vector
induction for use in fold-tail-with-masking, if a primary induction is absent.
The canonical scalar IV having start = 0 and step = VF*UF, created during code
-gen to control the vector loop, is widened into a canonical vector IV having
start = {<Part*VF, Part*VF+1, ..., Part*VF+VF-1> for 0 <= Part < UF} and
step = <VF*UF, VF*UF, ..., VF*UF>.
Differential Revision: https://reviews.llvm.org/D77635
When the Attributor was created the test update scripts were not well
suited to deal with the challenges of IR attribute checking. This
partially improved.
Since then we also added three additional configurations that need
testing; in total we now have the following four:
{ TUNIT, CGSCC } x { old pass manager (OPM), new pass manager (NPM) }
Finally, the number of developers and tests grew rapidly (partially due
to the addition of ArgumentPromotion and IPConstantProp tests), which
resulted in tests only being run in some configurations, different
prefixes being used, and different "styles" of checks being used.
Due to the above reasons I believed we needed to take another look at
the test update scripts. While we started to use them, via UTC_ARGS:
--enable/disable, the other problems remained. To improve the testing
situation for *all* configurations, to simplify future updates to the
test, and to help identify subtle effects of future changes, we now use
the test update scripts for (almost) all Attributor tests.
An exhaustive prefix list minimizes the number of check lines and makes
it easy to identify and compare configurations.
Tests have been adjusted in the process but we tried to keep their
intend unchanged.
Reviewed By: sstefan1
Differential Revision: https://reviews.llvm.org/D76588
Based on the post-commit comments for rG0f56bbc, there might
be a problem with this transform:
(bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
...and the ppc_fp128 data type, so conservatively bypass if we
are bitcasting a ppc_fp128.
We might be able to account for endian or other differences to
enable this for PowerPC again if that is useful.
Differential Revision: https://reviews.llvm.org/D77642
This patch updates the code that deals with conditions from predicate
info to make use of constant ranges.
For ssa_copy instructions inserted by PredicateInfo, we have 2 ranges:
1. The range of the original value.
2. The range imposed by the linked condition.
1. is known, 2. can be determined using makeAllowedICmpRegion. The
intersection of those ranges is the range for the copy.
With this patch, we get a nice increase in the number of instructions
eliminated by both SCCP and IPSCCP for some benchmarks:
For MultiSource, SPEC2000 & SPEC2006:
Tests: 237
Same hash: 170 (filtered out)
Remaining: 67
Metric: sccp.NumInstRemoved
Program base patch diff
test-suite...Source/Benchmarks/sim/sim.test 10.00 71.00 610.0%
test-suite...CFP2000/177.mesa/177.mesa.test 361.00 1626.00 350.4%
test-suite...encode/alacconvert-encode.test 141.00 602.00 327.0%
test-suite...decode/alacconvert-decode.test 141.00 602.00 327.0%
test-suite...CI_Purple/SMG2000/smg2000.test 1639.00 4093.00 149.7%
test-suite...peg2/mpeg2dec/mpeg2decode.test 75.00 163.00 117.3%
test-suite...T2006/401.bzip2/401.bzip2.test 358.00 513.00 43.3%
test-suite...rks/FreeBench/pifft/pifft.test 11.00 15.00 36.4%
test-suite...langs-C/unix-tbl/unix-tbl.test 4.00 5.00 25.0%
test-suite...lications/sqlite3/sqlite3.test 541.00 667.00 23.3%
test-suite.../CINT2000/254.gap/254.gap.test 243.00 299.00 23.0%
test-suite...ks/Prolangs-C/agrep/agrep.test 25.00 29.00 16.0%
test-suite...marks/7zip/7zip-benchmark.test 1135.00 1304.00 14.9%
test-suite...lications/ClamAV/clamscan.test 1105.00 1268.00 14.8%
test-suite...urce/Applications/lua/lua.test 398.00 436.00 9.5%
Metric: sccp.IPNumInstRemoved
Program base patch diff
test-suite...C/CFP2000/179.art/179.art.test 1.00 3.00 200.0%
test-suite...006/447.dealII/447.dealII.test 429.00 1056.00 146.2%
test-suite...nch/fourinarow/fourinarow.test 3.00 7.00 133.3%
test-suite...CI_Purple/SMG2000/smg2000.test 818.00 1748.00 113.7%
test-suite...ks/McCat/04-bisect/bisect.test 3.00 5.00 66.7%
test-suite...CFP2000/177.mesa/177.mesa.test 165.00 255.00 54.5%
test-suite...ediabench/gsm/toast/toast.test 18.00 27.00 50.0%
test-suite...telecomm-gsm/telecomm-gsm.test 18.00 27.00 50.0%
test-suite...ks/Prolangs-C/agrep/agrep.test 24.00 35.00 45.8%
test-suite...TimberWolfMC/timberwolfmc.test 43.00 62.00 44.2%
test-suite...encode/alacconvert-encode.test 46.00 66.00 43.5%
test-suite...decode/alacconvert-decode.test 46.00 66.00 43.5%
test-suite...langs-C/unix-tbl/unix-tbl.test 12.00 17.00 41.7%
test-suite...peg2/mpeg2dec/mpeg2decode.test 31.00 41.00 32.3%
test-suite.../CINT2000/254.gap/254.gap.test 117.00 154.00 31.6%
Reviewers: efriedma, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D76611
Ehud Katz