entire SCC before iterating on newly-introduced call edges resulting
from any inlined function bodies.
This more closely matches the behavior of the old PM's inliner. While it
wasn't really clear to me initially, this behavior is actually essential
to the inliner behaving reasonably in its current design.
Because the inliner is fundamentally a bottom-up inliner and all of its
cost modeling is designed around that it often runs into trouble within
an SCC where we don't have any meaningful bottom-up ordering to use. In
addition to potentially cyclic, infinite inlining that we block with the
inline history mechanism, it can also take seemingly simple call graph
patterns within an SCC and turn them into *insanely* large functions by
accidentally working top-down across the SCC without any of the
threshold limitations that traditional top-down inliners use.
Consider this diabolical monster.cpp file that Richard Smith came up
with to help demonstrate this issue:
```
template <int N> extern const char *str;
void g(const char *);
template <bool K, int N> void f(bool *B, bool *E) {
if (K)
g(str<N>);
if (B == E)
return;
if (*B)
f<true, N + 1>(B + 1, E);
else
f<false, N + 1>(B + 1, E);
}
template <> void f<false, MAX>(bool *B, bool *E) { return f<false, 0>(B, E); }
template <> void f<true, MAX>(bool *B, bool *E) { return f<true, 0>(B, E); }
extern bool *arr, *end;
void test() { f<false, 0>(arr, end); }
```
When compiled with '-DMAX=N' for various values of N, this will create an SCC
with a reasonably large number of functions. Previously, the inliner would try
to exhaust the inlining candidates in a single function before moving on. This,
unfortunately, turns it into a top-down inliner within the SCC. Because our
thresholds were never built for that, we will incrementally decide that it is
always worth inlining and proceed to flatten the entire SCC into that one
function.
What's worse, we'll then proceed to the next function, and do the exact same
thing except we'll skip the first function, and so on. And at each step, we'll
also make some of the constant factors larger, which is awesome.
The fix in this patch is the obvious one which makes the new PM's inliner use
the same technique used by the old PM: consider all the call edges across the
entire SCC before beginning to process call edges introduced by inlining. The
result of this is essentially to distribute the inlining across the SCC so that
every function incrementally grows toward the inline thresholds rather than
allowing the inliner to grow one of the functions vastly beyond the threshold.
The code for this is a bit awkward, but it works out OK.
We could consider in the future doing something more powerful here such as
prioritized order (via lowest cost and/or profile info) and/or a code-growth
budget per SCC. However, both of those would require really substantial work
both to design the system in a way that wouldn't break really useful
abstraction decomposition properties of the current inliner and to be tuned
across a reasonably diverse set of code and workloads. It also seems really
risky in many ways. I have only found a single real-world file that triggers
the bad behavior here and it is generated code that has a pretty pathological
pattern. I'm not worried about the inliner not doing an *awesome* job here as
long as it does *ok*. On the other hand, the cases that will be tricky to get
right in a prioritized scheme with a budget will be more common and idiomatic
for at least some frontends (C++ and Rust at least). So while these approaches
are still really interesting, I'm not in a huge rush to go after them. Staying
even closer to the existing PM's behavior, especially when this easy to do,
seems like the right short to medium term approach.
I don't really have a test case that makes sense yet... I'll try to find a
variant of the IR produced by the monster template metaprogram that is both
small enough to be sane and large enough to clearly show when we get this wrong
in the future. But I'm not confident this exists. And the behavior change here
*should* be unobservable without snooping on debug logging. So there isn't
really much to test.
The test case updates come from two incidental changes:
1) We now visit functions in an SCC in the opposite order. I don't think there
really is a "right" order here, so I just update the test cases.
2) We no longer compute some analyses when an SCC has no call instructions that
we consider for inlining.
llvm-svn: 297374
Summary:
In a .symver assembler directive like:
.symver name, name2@@nodename
"name2@@nodename" should get the same symbol binding as "name".
While the ELF object writer is updating the symbol binding for .symver
aliases before emitting the object file, not doing so when the module
inline assembly is handled by the RecordStreamer is causing the wrong
behavior in *LTO mode.
E.g. when "name" is global, "name2@@nodename" must also be marked as
global. Otherwise, the symbol is skipped when iterating over the LTO
InputFile symbols (InputFile::Symbol::shouldSkip). So, for example,
when performing any *LTO via the gold-plugin, the versioned symbol
definition is not recorded by the plugin and passed back to the
linker. If the object was in an archive, and there were no other symbols
needed from that object, the object would not be included in the final
link and references to the versioned symbol are undefined.
The llvm-lto2 tests added will give an error about an unused symbol
resolution without the fix.
Reviewers: rafael, pcc
Reviewed By: pcc
Subscribers: mehdi_amini, llvm-commits
Differential Revision: https://reviews.llvm.org/D30485
llvm-svn: 297332
!type metadata can not be dropped. An alternative to this is adding
!type metadata from the replaced globals to the replacement, but that
may weaken type tests and make them slower at the same time.
The merged global gets !dbg metadata from replaced globals, and can
end up with multiple debug locations.
llvm-svn: 297327
Analyzing larger trees is extremely difficult with the current debug output so
this adds GraphTraits and DOTGraphTraits on top of the VectorizableTree data
structure. We can now display the SLP trees with Graphviz as in
https://reviews.llvm.org/F3132765.
I decorated the graph where a value needs to be gathered for one reason or
another. These are the red nodes.
There are other improvement I am planning to make as I work through my case
here. For example, I would also like to mark nodes that need to be extracted.
Differential Revision: https://reviews.llvm.org/D30731
llvm-svn: 297303
Because IRBuilder performs constant-folding, it's not guaranteed that an
instruction in the original loop map to an instruction in the vector loop. It
could map to a constant vector instead. The handling of first-order recurrences
was incorrectly making this assumption when setting the IRBuilder's insert
point.
llvm-svn: 297302
Summary: Use AA when scanning to find an available load value.
Reviewers: rengolin, mcrosier, hfinkel, trentxintong, dberlin
Reviewed By: rengolin, dberlin
Subscribers: aemerson, dberlin, llvm-commits
Differential Revision: https://reviews.llvm.org/D30352
llvm-svn: 297284
Recommitting patch which was previously reverted in r297159. These
changes should address the casting issues.
The original patch enables dbg.value intrinsics to be attached to
newly inserted PHI nodes.
Differential Review: https://reviews.llvm.org/D30701
llvm-svn: 297269
Itanium ABI may have an address point one byte after the end of a
vtable. When such vtable global is split, the !type metadata needs to
follow the right vtable.
Differential Revision: https://reviews.llvm.org/D30716
llvm-svn: 297236
This was committed at r297155 and reverted at r297166 because of an
over-reaching clang test. That should be fixed with r297189.
This is one part of solving a recent bug report:
http://lists.llvm.org/pipermail/llvm-dev/2017-February/110293.html
This keeps with our general approach: changing arbitrary shuffles is off-limts,
but changing splat is ok. The transform is very similar to the existing
shrinkBitwiseLogic() canonicalization.
Differential Revision: https://reviews.llvm.org/D30123
llvm-svn: 297232
Summary:
The purpose of coro.end intrinsic is to allow frontends to mark the cleanup and
other code that is only relevant during the initial invocation of the coroutine
and should not be present in resume and destroy parts.
In landing pads coro.end is replaced with an appropriate instruction to unwind to
caller. The handling of coro.end differs depending on whether the target is
using landingpad or WinEH exception model.
For landingpad based exception model, it is expected that frontend uses the
`coro.end`_ intrinsic as follows:
```
ehcleanup:
%InResumePart = call i1 @llvm.coro.end(i8* null, i1 true)
br i1 %InResumePart, label %eh.resume, label %cleanup.cont
cleanup.cont:
; rest of the cleanup
eh.resume:
%exn = load i8*, i8** %exn.slot, align 8
%sel = load i32, i32* %ehselector.slot, align 4
%lpad.val = insertvalue { i8*, i32 } undef, i8* %exn, 0
%lpad.val29 = insertvalue { i8*, i32 } %lpad.val, i32 %sel, 1
resume { i8*, i32 } %lpad.val29
```
The `CoroSpit` pass replaces `coro.end` with ``True`` in the resume functions,
thus leading to immediate unwind to the caller, whereas in start function it
is replaced with ``False``, thus allowing to proceed to the rest of the cleanup
code that is only needed during initial invocation of the coroutine.
For Windows Exception handling model, a frontend should attach a funclet bundle
referring to an enclosing cleanuppad as follows:
```
ehcleanup:
%tok = cleanuppad within none []
%unused = call i1 @llvm.coro.end(i8* null, i1 true) [ "funclet"(token %tok) ]
cleanupret from %tok unwind label %RestOfTheCleanup
```
The `CoroSplit` pass, if the funclet bundle is present, will insert
``cleanupret from %tok unwind to caller`` before
the `coro.end`_ intrinsic and will remove the rest of the block.
Reviewers: majnemer
Reviewed By: majnemer
Subscribers: llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D25543
llvm-svn: 297223
When expanding the set of uniform instructions beyond the seed instructions
(e.g., consecutive pointers), we mark a new instruction uniform if all its
loop-varying users are uniform. We should also allow users that are consecutive
or interleaved memory accesses. This fixes cases where we have an instruction
that is used as the pointer operand of a consecutive access but also used by a
non-memory instruction that later becomes uniform as part of the expansion.
llvm-svn: 297179
This reverts commit r296488.
As noted by David Blaikie on llvm-commits, I overlooked the case of a
debug function being inlined into a nodebug function being inlined
into a debug function.
llvm-svn: 297163
Summary:
We should check if loop size allows us to peel at least one iteration
before we do so.
Patch by Max Kazantsev!
Reviewers: sanjoy, mkuper, efriedma
Reviewed By: mkuper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30632
llvm-svn: 297122
Summary: We do not need that special handling because the debug info is more accurate now. Performance testing shows no regression on google internal benchmarks.
Reviewers: davidxl, aprantl
Reviewed By: aprantl
Subscribers: llvm-commits, aprantl
Differential Revision: https://reviews.llvm.org/D30658
llvm-svn: 297038
LoopInfo::getLoopFor returns nullptr if a BB is not in a loop and only
then can the loop be updated to contain the newly created BBs. Add the
missing nullptr check to SplitBlockAndInsertIfThen.
Within LLVM, the only user of this function that also passes a LoopInfo
to be updated is InnerLoopVectorizer::predicateInstructions().
As the method's name implies, the BB operataten on will always be within
a loop, but out-of-tree users may also use it differently (here: Polly).
All other uses of LoopInfo::getLoopFor in the file properly check its
return value for nullptr.
llvm-svn: 297016
Any unsuccessful llvm.type.checked.load devirtualizations will be translated
into uses of llvm.type.test, so we need to add the resulting llvm.type.test
intrinsics to the function summaries so that the LowerTypeTests pass will
export them.
Differential Revision: https://reviews.llvm.org/D29808
llvm-svn: 296939
Summary:
If a loop contains a Phi node which has an invariant input from back
edge, it is profitable to peel such loops (rather than unroll them) to
use the advantage that this Phi is always invariant starting from 2nd
iteration. After the 1st iteration is peeled, other optimizations can
potentially simplify calculations with this invariant.
Patch by Max Kazantsev!
Reviewers: sanjoy, apilipenko, igor-laevsky, anna, mkuper, reames
Reviewed By: mkuper
Subscribers: mkuper, mzolotukhin, llvm-commits
Differential Revision: https://reviews.llvm.org/D30161
llvm-svn: 296898
Summary:
In current implementation the loop peeling happens after trip-count based partial unrolling and may
sometimes not happen at all due to it (for example, if trip count is known, but UP.Partial = false). This
is generally bad, the more than there are some situations where peeling is profitable even if the partial
unrolling is disabled.
This patch is a NFC which reorders peeling and partial unrolling application and prepares the code for
implementation of the said optimizations.
Patch by Max Kazantsev!
Reviewers: sanjoy, anna, reames, apilipenko, igor-laevsky, mkuper
Reviewed By: mkuper
Subscribers: mkuper, llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D30243
llvm-svn: 296897
for VectorizeTree() API.This API uses it for proper mask computation to be used in shufflevector IR.
The fix is to compute the mask for out of order memory accesses while building the vectorizable tree
instead of actual vectorization of vectorizable tree.It also needs to recompute the proper Lane for
external use of vectorizable scalars based on shuffle mask.
Reviewers: mkuper
Differential Revision: https://reviews.llvm.org/D30159
Change-Id: Ide8773ce0ad3562f3cf4d1a0ad0f487e2f60ce5d
llvm-svn: 296863
Such edges may otherwise result in infinite recursion if a pointer to a vtable
is reachable from the vtable itself. This can happen in practice if a TU
defines the ABI types used to implement RTTI, and is itself compiled with RTTI.
Fixes PR32121.
llvm-svn: 296839
ValueTracking is used for more thorough analysis of operands. Based on the
analysis, either run-time checks can be simplified (e.g. check only one operand
instead of two) or the transformation can be avoided. For example, it is quite
often the case that a divisor is promoted from a shorter type and run-time
checks for it are redundant.
With additional compile-time analysis of values, two special cases naturally
arise and are addressed by the patch:
1) Both operands are known to be short enough. Then, the long division can be
simply replaced with a short one without CFG modification.
2) If a division is unsigned and the dividend is known to be short then the
long division is not needed at all. Because if the divisor is too big for
short division then the quotient is obviously zero (and the remainder is
equal to the dividend). Actually, the division is not needed when
(divisor > dividend).
Differential Revision: https://reviews.llvm.org/D29897
llvm-svn: 296832
Chandler Carruth