This patch adds a couple of NewPM function passes (dot-dom and
dot-dom-only) that dump DomTree into .dot files.
Reviewed-By: aeubanks
Differential Revision: https://reviews.llvm.org/D116629
This reverts commit 640beb38e7.
That commit caused performance degradtion in Quicksilver test QS:sGPU and a functional test failure in (rocPRIM rocprim.device_segmented_radix_sort).
Reverting until we have a better solution to s_cselect_b64 codegen cleanup
Change-Id: Ibf8e397df94001f248fba609f072088a46abae08
Reviewed By: kzhuravl
Differential Revision: https://reviews.llvm.org/D115960
Change-Id: Id169459ce4dfffa857d5645a0af50b0063ce1105
The naming has come up as a source of confusion in several recent reviews. onlyWritesMemory is consist with onlyReadsMemory which we use for the corresponding readonly case as well.
In particular, this also preserves undef when loading from padding,
rather than converting it to zero through a different codepath.
This is the remaining part of D115924.
There are a number of places that specially handle loads from a
uniform value where all the bits are the same (zero, one, undef,
poison), because we a) don't care about the load offset in that
case b) it bypasses casts that might not be legal generally but
do work with uniform values.
We had multiple implementations of this, with a different set of
supported values each time. This replaces two usages with a more
complete helper. Other usages will be replaced separately, because
they have larger impact.
This is part of D115924.
This was noted in post-commit review for D116322 / 0edf99950e .
I am not seeing how to expose the bug in a test though because
we don't pass an assumption cache into this analysis from there.
This ports the logic we generate in instcombine for a single use x.with.overflow check for use in SCEV's analysis. The result is that we can prove trip counts for many checks, and (through existing logic) often discharge them.
Motivation comes from compiling a simple example with -ftrapv.
Differential Revision: https://reviews.llvm.org/D116499
0a00d64 turned an early exit here into an assertion, but the assertion
can be triggered, as PR52920 shows.
The later code is agnostic to the accessed type, so just drop the
assert. The patch also adds tests for LAA directly and
loop-load-elimination to show the behavior is sane.
For loops that contain in-loop reductions but no loads or stores, large
VFs are chosen because LoopVectorizationCostModel::getSmallestAndWidestTypes
has no element types to check through and so returns the default widths
(-1U for the smallest and 8 for the widest). This results in the widest
VF being chosen for the following example,
float s = 0;
for (int i = 0; i < N; ++i)
s += (float) i*i;
which, for more computationally intensive loops, leads to large loop
sizes when the operations end up being scalarized.
In this patch, for the case where ElementTypesInLoop is empty, the widest
type is determined by finding the smallest type used by recurrences in
the loop instead of falling back to a default value of 8 bits. This
results in the cost model choosing a more sensible VF for loops like
the one above.
Differential Revision: https://reviews.llvm.org/D113973
This function returns an upper bound on the number of bits needed
to represent the signed value. Use "Max" to match similar functions
in KnownBits like countMaxActiveBits.
Rename APInt::getMinSignedBits->getSignificantBits. Keeping the old
name around to keep this patch size down. Will do a bulk rename as
follow up.
Rename KnownBits::countMaxSignedBits->countMaxSignificantBits.
Reviewed By: lebedev.ri, RKSimon, spatel
Differential Revision: https://reviews.llvm.org/D116522
This reverts commit fd4808887e.
This patch causes gcc to issue a lot of warnings like:
warning: base class ‘class llvm::MCParsedAsmOperand’ should be
explicitly initialized in the copy constructor [-Wextra]
We can fold an equality or unsigned icmp between base+offset1 and
base+offset2 with inbounds offsets by comparing the offsets directly.
This replaces a pair of specialized folds that tried to reason
based on the GEP structure instead. One of those folds was plain
wrong (because it does not account for negative offsets), while
the other is unnecessarily complicated and limited (e.g. it will
fail with bitcasts involved).
The disadvantage of this change is that it requires data layout,
so the fold is no longer performed by datalayout-independent
constant folding. I don't think this is a loss in practice, but
it does regress the ConstantExprFold.ll test, which checks folding
without running any passes.
Differential Revision: https://reviews.llvm.org/D116332
We should not lose analysis precision if an 'add' has both no-wrap
flags (nsw and nuw) compared to just one or the other.
This patch is modeled on a similar construct that was added with
D59386.
I don't think it is possible to expose a problem with an unsigned
compare because of the way this was coded (nuw is handled first).
InstCombine has an assert that fires with the example from:
https://github.com/llvm/llvm-project/issues/52884
...because it was expecting InstSimplify to handle this kind of
pattern with an smax.
Fixes#52884
Differential Revision: https://reviews.llvm.org/D116322
Remove the special casing for intrinsics in MemoryLocation::getForDest()
and handle them through the general attribute based code. On the DSE
side, this means that isRemovable() now needs to handle more than a
hardcoded list of intrinsics. We consider everything apart from
volatile memory intrinsics and lifetime markers to be removable.
This allows us to perform DSE on intrinsics that DSE has not been
specially taught about, using a matrix store as an example here.
There is an interesting test change for invariant.start, but I
believe that optimization is correct. It only looks a bit odd
because the code is immediate UB anyway.
Differential Revision: https://reviews.llvm.org/D116210
Adding following fold opportunity:
((A | B) ^ A) & ((A | B) ^ B) --> 0
Reviewed By: spatel, rampitec
Differential Revision: https://reviews.llvm.org/D115755
When looking at building the generator for regalloc, we realized we'd
need quite a bit of custom logic, and that perhaps it'd be easier to
just have each usecase (each kind of mlgo policy) have it's own
stand-alone test generator.
This patch just consolidates the old `config.py` and
`generate_mock_model.py` into one file, and does away with
subdirectories under Analysis/models.
As reames mentioned on related reviews, we don't need the nocapture
requirement here. First of all, from an API perspective, this is
not something that MemoryLocation::getForDest() should be checking
in the first place, because it does not affect which memory this
particular call can access; it's an orthogonal concern that should
be handled by the caller if necessary.
However, for both of the motivating users in DSE and InstCombine,
we don't need the nocapture requirement, because the capture can
either be purely local to the call (a pointer identity check that
is irrelevant to us), be part of the return value (which we check
is unused), or be written in the dest location, which we have
determined to be dead.
This allows us to remove the special handling for libcalls as well.
Differential Revision: https://reviews.llvm.org/D116148
The areFunctionArgsABICompatible() hook currently accepts a list of
pointer arguments, though what we're actually interested in is the
ABI compatibility after these pointer arguments have been converted
into value arguments.
This means that a) the current API is incompatible with opaque
pointers (because it requires inspection of pointee types) and
b) it can only be used in the specific context of ArgPromotion.
I would like to reuse the API when inspecting calls during inlining.
This patch converts it into an areTypesABICompatible() hook, which
accepts a list of types. This makes the method more generally usable,
and compatible with opaque pointers from an API perspective (the
actual usage in ArgPromotion/Attributor is still incompatible,
I'll follow up on that in separate patches).
Differential Revision: https://reviews.llvm.org/D116031
This fixes the assertion failure reported at
https://reviews.llvm.org/D114889#3198921 with a straightforward
check, until the cleaner fix in D115924 can be reapplied.
This is a reapply of a8a51fe5, which was reverted in 1ba99e due to a failing compiler-rt test. That test was a false positive because it was checking asan failures not accounting for the fact the call could be validly optimized out. I hopefully managed to stablize that test in 9b955f. (That's a speculative fix due to disk consumption needed to build compiler-rt tests locally being absurd.)
Original commit message follows..
The majority of this change is sinking logic from instcombine into MemoryLocation such that it can be generically reused. If we have a call with a single analyzable write to an argument, we can treat that as-if it were a store of unknown size.
Merging the code in this was unblocks DSE in the store to dead memory code paths. In theory, it should also enable classic DSE of such calls, but the code appears to not know how to use object sizes to refine unknown access bounds (yet).
In addition, this does make the isAllocRemovable path slightly stronger by reusing the libfunc and additional intrinsics bits which are already in getForDest.
Differential Revision: https://reviews.llvm.org/D115904
The availability of SVE should be sufficient to enable scalable
auto-vectorization.
This patch adds a new TTI interface to query the target what style of
vectorization it wants when scalable vectors are available. For other
targets than AArch64, this currently defaults to 'FixedWidthOnly'.
Differential Revision: https://reviews.llvm.org/D115651
This reverts commit 9fd4f80e33.
This breaks SingleSource/Regression/C/gcc-c-torture/execute/pr19687.c
in test-suite. Either the test is incorrect, or clang is generating
incorrect union initialization code. I've submitted
https://reviews.llvm.org/D115994 to fix the test, assuming my
interpretation is correct. Reverting this in the meantime as it
may take some time to resolve.
Before this change, AAResults::getModRefInfo() was missing a case for
callbr instructions (asm goto), which may read/write memory. In PR52735,
this led to a miscompile where a load was incorrect eliminated.
Add this missing case, as well as an assert verifying that all
memory-accessing instructions are handled properly.
Fixes#52735.
Differential Revision: https://reviews.llvm.org/D115992
The majority of this change is sinking logic from instcombine into MemoryLocation such that it can be generically reused. If we have a call with a single analyzable write to an argument, we can treat that as-if it were a store of unknown size.
Merging the code in this was unblocks DSE in the store to dead memory code paths. In theory, it should also enable classic DSE of such calls, but the code appears to not know how to use object sizes to refine unknown access bounds (yet).
In addition, this does make the isAllocRemovable path slightly stronger by reusing the libfunc and additional intrinsics bits which are already in getForDest.
Differential Revision: https://reviews.llvm.org/D115904
Pull out an explicit check rather than relying on the fact that the callee operand is not a data operand. The only real value is it gives us a clear place to move the comment, and makes the code slightly more understandable.
There are a number of places that specially handle loads from a
uniform value where all the bits are the same (zero, one, undef,
poison), because we a) don't care about the load offset in that
case and b) it bypasses casts that might not be legal generally
but do work with uniform values.
We had multiple implementations of this, with a different set of
supported values each time, as well as incomplete type checks in
some cases. In particular, this fixes the assertion reported in
https://reviews.llvm.org/D114889#3198921, as well as a similar
assertion that could be triggered via constant folding.
Differential Revision: https://reviews.llvm.org/D115924
Preserve the invariant that offset reported in the case of a
`PartialAlias` between `Loc1` and `Loc2`, is such that
`Loc1 + Offset = Loc2`, where `Loc1` and `Loc2` are the first and
the second argument, respectively, in alias queries.
Differential Revision: https://reviews.llvm.org/D115927
This patch updates applyLoopGuards to first collect all conditions and
then applies them in reverse order. This ensures the SCEVs with the
shortest dependency chains are constructed first, limiting the required
stack size.
This fixes a crash reported in D113578.
Note that the order conditions are applied can impact the accuracy of
the result, mostly due to missing min/max simplifications when
constructing SCEVs.
The changed test highlights the impact of the evaluation order. I will
follow up with a SCEV patch to improve min/max simplifications to get
the same results for both orders.
After the switch to the new pass manager, we have observed multiple
instances of catastrophic inlining, where the inliner produces huge
functions with many hundreds of thousands of instructions from small
input IR. We were forced to back out the switch to the new pass
manager for this reason. This patch fixes at least one of the root
cause issues.
LLVM uses a bottom-up inliner, and the fact that functions are processed
bottom-up is not just a question of optimality -- it is an imporant
requirement to prevent runaway inlining. The premise of the current
inlining approach and cost model is that after all calls inside a function
have been inlined, it may get large enough that inlining it into its
callers is no longer considered profitable. This safeguard does not
exist if inlining doesn't happen bottom-up, as inlining the callees,
and their callees, and their callees etc. will always seem individually
profitable, and the inliner can easily flatten the whole call tree.
There are instances where we necessarily have to deviate from bottom-up
inlining: When inlining in an SCC there is no natural "bottom", so
inlining effectively happens top-down. This requires special care,
and the inliner avoids exponential blowup by ensuring that functions
in the SCC grow in a balanced way and will eventually hit the threshold.
However, there is one instance where the inlining advisor explicitly
violates the bottom-up principle: Deferred inlining tries to "defer"
inlining a call if it determines that inlining the caller into all
its call-sites would be more profitable. Something very important to
understand about deferred inlining is that it doesn't make one inlining
choice in place of another -- it effectively chooses to do both. If we
have a call chain A -> B -> C and cost modelling tells us that inlining
B -> C is profitable, but we defer this and instead inline A -> B first,
then we'll now have a call A -> C, and the cost model will (a few special
cases notwithstanding) still tell us that this is profitable. So the end
result is that we inlined *both* B and C, even though under the usual
cost model function B would have been too large to further inline after
C has been integrated into it.
Because deferred inlining violates the bottom-up invariant of the inliner,
it can result in exponential inlining. The exponential-deferred-inlining.ll
test case illustrates this on a simple example (see
https://gist.github.com/nikic/1262b5f7d27278e1b34a190ae10947f5 for a
much more catastrophic case with about 5000x size blowup). If the call
chain A -> B -> C is not a chain but a tree of calls, then we end up
deferring inlining across the tree and end up flattening everything into
the root node.
This patch proposes to address this by disabling deferred inlining
entirely (currently still behind an option). Beyond the issue of
exponential inlining, I don't think that the whole concept makes sense,
at least as long as deferred inlining still ends up inlining both call
edges.
I believe the motivation for having deferred inlining in the first place
is that you might have a small wrapper function with local linkage that
could be eliminated if inlined. This would automatically happen if there
was a single caller, due to the large "last call to local" bonus. However,
this bonus is not extended if there are multiple callers, even if we
would eventually end up inlining into all of them (if the bonus were
extended).
Now, unlike the normal inlining cost model, the deferred inlining cost
model does look at all callers, and will extend the "last call to local"
bonus if it determines that we could inline all of them as long as we
defer the current inlining decision. This makes very little sense.
The "last call to local" bonus doesn't really cost model anything.
It's basically an "infinite" bonus that ensures we always inline the
last call to a local. The fact that it's not literally infinite just
prevents inlining of huge functions, which can easily result in
scalability issues. I very much doubt that it was an intentional
cost-modelling choice to say that getting rid of a small local function
is worth adding 15000 instructions elsewhere, yet this is exactly how
this value is getting used here.
The main alternative I see to complete removal is to change deferred
inlining to an actual either/or decision. That is, to mark deferred
calls as noinline so we're actually trading off one inlining decision
against another, and not just adding a side-channel to the cost model
to do both.
Apart from fixing the catastrophic inlining case, the effect on rustc
is a modest compile-time improvement on average (up to 8% for a
parsing-type crate, where tree-like calls are expected) and pretty
neutral where run-time performance is concerned (mix of small wins
and losses, usually in the sub-1% category).
Differential Revision: https://reviews.llvm.org/D115497
