This change leverages the work done in D83743 to replay in the SampleProfile inliner to also be used in the CGSCC inliner. NOTE: currently restricted to non-ML advisors only.
The added switch `-cgscc-inline-replay=<remarks file>` will replay the inlining decisions in that file where the remarks file is generated via `-Rpass=inline`. The aim here is to make it easier to analyze changes that would modify inlining heuristics to be separated from this behavior. Doing so allows easier examination of assembly and runtime behavior compared to the baseline rather than trying to dig through the large churn caused by inlining.
In LTO compilation, since inlining is done twice you can separately specify replay by passing the flag to the FE (`-cgscc-inline-replay=`) and to the linker (`-Wl,cgscc-inline-replay=`) with the remarks generated from their respective places.
Testing on mysqld by comparing the inline decisions between base (generates remarks.txt) and diff (replay using identical input/tools with remarks.txt) and examining the inlining sites with `diff` shows 14,000 mismatches out of 247,341 for a ~94% replay accuracy. I believe this gap can be narrowed further though for the general case we may never achieve full accuracy. For my personal use, this is close enough to be representative: I set the baseline as the one generated by the replay on identical input/toolset and compare that to my modified input/toolset using the same replay.
Testing:
ninja check-llvm
newly added test correctly replays CGSCC inlining decisions
Reviewed By: mtrofin, wenlei
Differential Revision: https://reviews.llvm.org/D94334
This is similar to D94106, but for the
isGuaranteedToTransferExecutionToSuccessor() helper. We should not
assume that readonly functions will return, as this is only true for
mustprogress functions (in which case we already infer willreturn).
As with the DCE change, for now continue assuming that readonly
intrinsics will return, as not all target intrinsics have been
annotated yet.
Differential Revision: https://reviews.llvm.org/D95288
This is to support the memory routines vec_malloc, vec_calloc, vec_realloc, and vec_free. These routines manage memory that is 16-byte aligned. And they are only available on AIX.
Differential Revision: https://reviews.llvm.org/D94710
We tend to assume that the AA pipeline is by default the default AA
pipeline and it's confusing when it's empty instead.
PR48779
Initially reverted due to BasicAA running analyses in an unspecified
order (multiple function calls as parameters), fixed by fetching
analyses before the call to construct BasicAA.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D95117
Having a custom inliner doesn't really fit in with the new PM's
pipeline. It's also extra technical debt.
amdgpu-inline only does a couple of custom things compared to the normal
inliner:
1) It disables inlining if the number of BBs in a function would exceed
some limit
2) It increases the threshold if there are pointers to private arrays(?)
These can all be handled as TTI inliner hooks.
There already exists a hook for backends to multiply the inlining
threshold.
This way we can remove the custom amdgpu-inline pass.
This caused inline-hint.ll to fail, and after some investigation, it
looks like getInliningThresholdMultiplier() was previously getting
applied twice in amdgpu-inline (https://reviews.llvm.org/D62707 fixed it
not applying at all, so some later inliner change must have fixed
something), so I had to change the threshold in the test.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D94153
This adds cost modelling for the inloop vectorization added in
745bf6cf44. Up until now they have been modelled as the original
underlying instruction, usually an add. This happens to works OK for MVE
with instructions that are reducing into the same type as they are
working on. But MVE's instructions can perform the equivalent of an
extended MLA as a single instruction:
%sa = sext <16 x i8> A to <16 x i32>
%sb = sext <16 x i8> B to <16 x i32>
%m = mul <16 x i32> %sa, %sb
%r = vecreduce.add(%m)
->
R = VMLADAV A, B
There are other instructions for performing add reductions of
v4i32/v8i16/v16i8 into i32 (VADDV), for doing the same with v4i32->i64
(VADDLV) and for performing a v4i32/v8i16 MLA into an i64 (VMLALDAV).
The i64 are particularly interesting as there are no native i64 add/mul
instructions, leading to the i64 add and mul naturally getting very
high costs.
Also worth mentioning, under NEON there is the concept of a sdot/udot
instruction which performs a partial reduction from a v16i8 to a v4i32.
They extend and mul/sum the first four elements from the inputs into the
first element of the output, repeating for each of the four output
lanes. They could possibly be represented in the same way as above in
llvm, so long as a vecreduce.add could perform a partial reduction. The
vectorizer would then produce a combination of in and outer loop
reductions to efficiently use the sdot and udot instructions. Although
this patch does not do that yet, it does suggest that separating the
input reduction type from the produced result type is a useful concept
to model. It also shows that a MLA reduction as a single instruction is
fairly common.
This patch attempt to improve the costmodelling of in-loop reductions
by:
- Adding some pattern matching in the loop vectorizer cost model to
match extended reduction patterns that are optionally extended and/or
MLA patterns. This marks the cost of the reduction instruction correctly
and the sext/zext/mul leading up to it as free, which is otherwise
difficult to tell and may get a very high cost. (In the long run this
can hopefully be replaced by vplan producing a single node and costing
it correctly, but that is not yet something that vplan can do).
- getExtendedAddReductionCost is added to query the cost of these
extended reduction patterns.
- Expanded the ARM costs to account for these expanded sizes, which is a
fairly simple change in itself.
- Some minor alterations to allow inloop reduction larger than the highest
vector width and i64 MVE reductions.
- An extra InLoopReductionImmediateChains map was added to the vectorizer
for it to efficiently detect which instructions are reductions in the
cost model.
- The tests have some updates to show what I believe is optimal
vectorization and where we are now.
Put together this can greatly improve performance for reduction loop
under MVE.
Differential Revision: https://reviews.llvm.org/D93476
This reverts commit d97f776be5.
The original problem was due to build failures in shared lib builds. D95079
moved ImportedFunctionsInliningStatistics under Analysis, unblocking
this.
This is related to D94982. We want to call these APIs from the Analysis
component, so we can't leave them under Transforms.
Differential Revision: https://reviews.llvm.org/D95079
When using 2 InlinePass instances in the same CGSCC - one for other
mandatory inlinings, the other for the heuristic-driven ones - the order
in which the ImportedFunctionStats would be output-ed would depend on
the destruction order of the inline passes, which is not deterministic.
This patch moves the ImportedFunctionStats responsibility to the
InlineAdvisor to address this problem.
Differential Revision: https://reviews.llvm.org/D94982
Currently LLVM is relying on ValueTracking's `isKnownNonZero` to attach `nonnull`, which can return true when the value is poison.
To make the semantics of `nonnull` consistent with the behavior of `isKnownNonZero`, this makes the semantics of `nonnull` to accept poison, and return poison if the input pointer isn't null.
This makes many transformations like below legal:
```
%p = gep inbounds %x, 1 ; % p is non-null pointer or poison
call void @f(%p) ; instcombine converts this to call void @f(nonnull %p)
```
Instead, this semantics makes propagation of `nonnull` to caller illegal.
The reason is that, passing poison to `nonnull` does not immediately raise UB anymore, so such program is still well defined, if the callee does not use the argument.
Having `noundef` attribute there re-allows this.
```
define void @f(i8* %p) { ; functionattr cannot mark %p nonnull here anymore
call void @g(i8* nonnull %p) ; .. because @g never raises UB if it never uses %p.
ret void
}
```
Another attribute that needs to be updated is `align`. This patch updates the semantics of align to accept poison as well.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D90529
Just like llvm.assume, there are a lot of cases where we can just ignore llvm.experimental.noalias.scope.decl.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D93042
Split impliesPoison into two recursive walks, one over V, the
other over ValAssumedPoison. This allows us to reason about poison
implications in a number of additional cases that are important
in practice. This is a generalized form of D94859, which handles
the cmp to cmp implication in particular.
Differential Revision: https://reviews.llvm.org/D94866
83daa49758 made loop-rotate more conservative in the presence of
function calls in the prepare-for-lto stage. The code did not properly
account for calls that are no actual function calls, like calls to
intrinsics. This patch updates the code to ensure only calls that are
lowered to actual calls are considered inline candidates.
D84108 exposed a bad interaction between inlining and loop-rotation
during regular LTO, which is causing notable regressions in at least
CINT2006/473.astar.
The problem boils down to: we now rotate a loop just before the vectorizer
which requires duplicating a function call in the preheader when compiling
the individual files ('prepare for LTO'). But this then prevents further
inlining of the function during LTO.
This patch tries to resolve this issue by making LoopRotate more
conservative with respect to rotating loops that have inline-able calls
during the 'prepare for LTO' stage.
I think this change intuitively improves the current situation in
general. Loop-rotate tries hard to avoid creating headers that are 'too
big'. At the moment, it assumes all inlining already happened and the
cost of duplicating a call is equal to just doing the call. But with LTO,
inlining also happens during full LTO and it is possible that a previously
duplicated call is actually a huge function which gets inlined
during LTO.
From the perspective of LV, not much should change overall. Most loops
calling user-provided functions won't get vectorized to start with
(unless we can infer that the function does not touch memory, has no
other side effects). If we do not inline the 'inline-able' call during
the LTO stage, we merely delayed loop-rotation & vectorization. If we
inline during LTO, chances should be very high that the inlined code is
itself vectorizable or the user call was not vectorizable to start with.
There could of course be scenarios where we inline a sufficiently large
function with code not profitable to vectorize, which would have be
vectorized earlier (by scalarzing the call). But even in that case,
there probably is no big performance impact, because it should be mostly
down to the cost-model to reject vectorization in that case. And then
the version with scalarized calls should also not be beneficial. In a way,
LV should have strictly more information after inlining and make more
accurate decisions (barring cost-model issues).
There is of course plenty of room for things to go wrong unexpectedly,
so we need to keep a close look at actual performance and address any
follow-up issues.
I took a look at the impact on statistics for
MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer
loops rotated, but no change to the number of loops vectorized.
Reviewed By: sanwou01
Differential Revision: https://reviews.llvm.org/D94232
We can fold x*C1/C2 <= x to true if C1 <= C2. This is valid even
if the multiplication is not nuw: https://alive2.llvm.org/ce/z/vULors
The multiplication or division can be replaced by shifts. We don't
handle the case where both are shifts, as that should get folded
away by InstCombine.
There are no changes relative to the original commit. However, an issue
this exposed in BasicAA assumption tracking has been fixed in the
previous commit.
-----
An alias query currently works out roughly like this:
* Look up location pair in cache.
* Perform BasicAA logic (including cache lookup and insertion...)
* Perform a recursive query using BestAAResults.
* Look up location pair in cache (and thus do not recurse into BasicAA)
* Query all the other AA providers.
* Query all the other AA providers.
This is a lot of unnecessary work, all ultimately caused by the
BestAAResults query at the end of aliasCheck(). The reason we perform
it, is that aliasCheck() is getting called recursively, and we of
course want those recursive queries to also make use of other AA
providers, not just BasicAA. We can solve this by making the recursive
queries directly use BestAAResults (which will check both BasicAA
and other providers), rather than recursing into aliasCheck().
There are some tradeoffs:
* We can no longer pass through the precomputed underlying object
to aliasCheck(). This is not a major concern, because nowadays
getUnderlyingObject() is quite cheap.
* Results from other AA providers are no longer cached inside
BasicAA. The way this worked was already a bit iffy, in that a
result could be cached, but if it was MayAlias, we'd still end
up re-querying other providers anyway. If we want to cache
non-BasicAA results, we should do that in a more principled manner.
In any case, despite those tradeoffs, this works out to be a decent
compile-time improvment. I think it also simplifies the mental model
of how BasicAA works. It took me quite a while to fully understand
how these things interact.
Differential Revision: https://reviews.llvm.org/D90094
D91936 placed the tracking for the assumptions into BasicAA.
However, when recursing over phis, we may use fresh AAQI instances.
In this case AssumptionBasedResults from an inner AAQI can reesult
in a removal of an element from the outer AAQI.
To avoid this, move the tracking into AAQI. This generally makes
more sense, as the NoAlias assumptions themselves are also stored
in AAQI.
The test case only produces an assertion failure with D90094
reapplied. I think the issue exists independently of that change
as well, but I wasn't able to come up with a reproducer.
Expanding from D94808 - we ensure the same InlineAdvisor is used by both
InlinerPass instances. The notion of mandatory inlining is moved into
the core InlineAdvisor: advisors anyway have to handle that case, so
this change also factors out that a bit better.
Differential Revision: https://reviews.llvm.org/D94825
This reverts commit a3904cc77f.
It causes the compiler to crash while building Harfbuzz for ARM in
Chromium, reduced reproducer forthcoming:
https://crbug.com/1167305
An alias query currently works out roughly like this:
* Look up location pair in cache.
* Perform BasicAA logic (including cache lookup and insertion...)
* Perform a recursive query using BestAAResults.
* Look up location pair in cache (and thus do not recurse into BasicAA)
* Query all the other AA providers.
* Query all the other AA providers.
This is a lot of unnecessary work, all ultimately caused by the
BestAAResults query at the end of aliasCheck(). The reason we perform
it, is that aliasCheck() is getting called recursively, and we of
course want those recursive queries to also make use of other AA
providers, not just BasicAA. We can solve this by making the recursive
queries directly use BestAAResults (which will check both BasicAA
and other providers), rather than recursing into aliasCheck().
There are some tradeoffs:
* We can no longer pass through the precomputed underlying object
to aliasCheck(). This is not a major concern, because nowadays
getUnderlyingObject() is quite cheap.
* Results from other AA providers are no longer cached inside
BasicAA. The way this worked was already a bit iffy, in that a
result could be cached, but if it was MayAlias, we'd still end
up re-querying other providers anyway. If we want to cache
non-BasicAA results, we should do that in a more principled manner.
In any case, despite those tradeoffs, this works out to be a decent
compile-time improvment. I think it also simplifies the mental model
of how BasicAA works. It took me quite a while to fully understand
how these things interact.
Differential Revision: https://reviews.llvm.org/D90094
This change modifies the source location formatting from:
LineNumber.Discriminator
to:
LineNumber:ColumnNumber.Discriminator
The motivation here is to enhance location information for inline replay that currently exists for the SampleProfile inliner. This will be leveraged further in inline replay for the CGSCC inliner in the related diff.
The ReplayInlineAdvisor is also modified to read the new format and now takes into account the callee for greater accuracy.
Testing:
ninja check-llvm
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D94333
This is a partial fix for https://bugs.llvm.org/show_bug.cgi?id=44403.
Folding gep p, q-p to q is only legal if p and q have the same
provenance. This fold should probably be guarded by something like
getUnderlyingObject(p) == getUnderlyingObject(q).
This patch is a partial fix that removes the special handling for
gep p, 0-p, which will fold to a null pointer, which would certainly
not pass an underlying object check (unless p is also null, in which
case this would fold trivially anyway). Folding to a null pointer
is particularly problematic due to the special handling it receives
in many places, making end-to-end miscompiles more likely.
Differential Revision: https://reviews.llvm.org/D93820
This patch fixes a bug that could result in miscompiles (at least
in an OOT target). The problem could be seen by adding checks that
the DominatorTree used in BasicAliasAnalysis and ValueTracking was
valid (e.g. by adding DT->verify() call before every DT dereference
and then running all tests in test/CodeGen).
Problem was that the LegacyPassManager calculated "last user"
incorrectly for passes such as the DominatorTree when not telling
the pass manager that there was a transitive dependency between
the different analyses. And then it could happen that an incorrect
dominator tree was used when doing alias analysis (which was a pretty
serious bug as the alias analysis result could be invalid).
Fixes: https://bugs.llvm.org/show_bug.cgi?id=48709
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D94138
Currently make_early_inc_range cannot be used with iterators with
operator* implementations that do not return a reference.
Most notably in the LLVM codebase, this means the User iterator ranges
cannot be used with make_early_inc_range, which slightly simplifies
iterating over ranges while elements are removed.
Instead of directly using BaseT::reference as return type of operator*,
this patch uses decltype to get the actual return type of the operator*
implementation in WrappedIteratorT.
This patch also updates a few places to use make use of
make_early_inc_range.
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D93992
modimo