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
Previously, any change in any function in an SCC would cause all
analyses for all functions in the SCC to be invalidated. With this
change, we now manually invalidate analyses for functions we modify,
then let the pass manager know that all function analyses should be
preserved since we've already handled function analysis invalidation.
So far this only touches the inliner, argpromotion, function-attrs, and
updateCGAndAnalysisManager(), since they are the most used.
This is part of an effort to investigate running the function
simplification pipeline less on functions we visit multiple times in the
inliner pipeline.
However, this causes major memory regressions especially on larger IR.
To counteract this, turn on the option to eagerly invalidate function
analyses. This invalidates analyses on functions immediately after
they're processed in a module or scc to function adaptor for specific
parts of the pipeline.
Within an SCC, if a pass only modifies one function, other functions in
the SCC do not have their analyses invalidated, so in later function
passes in the SCC pass manager the analyses may still be cached. It is
only after the function passes that the eager invalidation takes effect.
For the default pipelines this makes sense because the inliner pipeline
runs the function simplification pipeline after all other SCC passes
(except CoroSplit which doesn't request any analyses).
Overall this has mostly positive effects on compile time and positive effects on memory usage.
https://llvm-compile-time-tracker.com/compare.php?from=7f627596977624730f9298a1b69883af1555765e&to=39e824e0d3ca8a517502f13032dfa67304841c90&stat=instructionshttps://llvm-compile-time-tracker.com/compare.php?from=7f627596977624730f9298a1b69883af1555765e&to=39e824e0d3ca8a517502f13032dfa67304841c90&stat=max-rss
D113196 shows that we slightly regressed compile times in exchange for
some memory improvements when turning on eager invalidation. D100917
shows that we slightly improved compile times in exchange for major
memory regressions in some cases when invalidating less in SCC passes.
Turning these on at the same time keeps the memory improvements while
keeping compile times neutral/slightly positive.
Reviewed By: asbirlea, nikic
Differential Revision: https://reviews.llvm.org/D113304
Printing pass manager invocations is fairly verbose and not super
useful.
This allows us to remove DebugLogging from pass managers and PassBuilder
since all logging (aside from analysis managers) goes through
instrumentation now.
This has the downside of never being able to print the top level pass
manager via instrumentation, but that seems like a minor downside.
Reviewed By: ychen
Differential Revision: https://reviews.llvm.org/D101797
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
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
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D95117
```
// The legacy PM CGPassManager discovers SCCs this way:
for function in the source order
tarjanSCC(function)
// While the new PM CGSCCPassManager does:
for function in the reversed source order [1]
discover a reference graph SCC
build call graph SCCs inside the reference graph SCC
```
In the common cases, reference graph ~= call graph, the new PM order is
undesired because for `a | b | c` (3 independent functions), the new PM will
process them in the reversed order: c, b, a. If `a <-> b <-> c`, we can see
that `-print-after-all` will report the sole SCC as `scc: (c, b, a)`.
This patch corrects the iteration order. The discovered SCC order will match
the legacy PM in the common cases.
For some tests (`Transforms/Inline/cgscc-*.ll` and
`unittests/Analysis/CGSCCPassManagerTest.cpp`), the behaviors are dependent on
the SCC discovery order and there are too many check lines for the particular
order. This patch simply reverses the function order to avoid changing too many
check lines.
Differential Revision: https://reviews.llvm.org/D90566
If an analysis is actually invalidated, there's already a log statement
for that: 'Invalidating analysis: FooAnalysis'.
Otherwise the statement is not very useful.
Reviewed By: asbirlea, ychen
Differential Revision: https://reviews.llvm.org/D84981
As it's causing some bot failures (and per request from kbarton).
This reverts commit r358543/ab70da07286e618016e78247e4a24fcb84077fda.
llvm-svn: 358546
The issue here is that we actually allow CGSCC passes to mutate IR (and
therefore invalidate analyses) outside of the current SCC. At a minimum,
we need to support mutating parent and ancestor SCCs to support the
ArgumentPromotion pass which rewrites all calls to a function.
However, the analysis invalidation infrastructure is heavily based
around not needing to invalidate the same IR-unit at multiple levels.
With Loop passes for example, they don't invalidate other Loops. So we
need to customize how we handle CGSCC invalidation. Doing this without
gratuitously re-running analyses is even harder. I've avoided most of
these by using an out-of-band preserved set to accumulate the cross-SCC
invalidation, but it still isn't perfect in the case of re-visiting the
same SCC repeatedly *but* it coming off the worklist. Unclear how
important this use case really is, but I wanted to call it out.
Another wrinkle is that in order for this to successfully propagate to
function analyses, we have to make sure we have a proxy from the SCC to
the Function level. That requires pre-creating the necessary proxy.
The motivating test case now works cleanly and is added for
ArgumentPromotion.
Thanks for the review from Philip and Wei!
Differential Revision: https://reviews.llvm.org/D59869
llvm-svn: 357137
Pass Execution Instrumentation interface enables customizable instrumentation
of pass execution, as per "RFC: Pass Execution Instrumentation interface"
posted 06/07/2018 on llvm-dev@
The intent is to provide a common machinery to implement all
the pass-execution-debugging features like print-before/after,
opt-bisect, time-passes etc.
Here we get a basic implementation consisting of:
* PassInstrumentationCallbacks class that handles registration of callbacks
and access to them.
* PassInstrumentation class that handles instrumentation-point interfaces
that call into PassInstrumentationCallbacks.
* Callbacks accept StringRef which is just a name of the Pass right now.
There were some ideas to pass an opaque wrapper for the pointer to pass instance,
however it appears that pointer does not actually identify the instance
(adaptors and managers might have the same address with the pass they govern).
Hence it was decided to go simple for now and then later decide on what the proper
mental model of identifying a "pass in a phase of pipeline" is.
* Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies
on different IRUnits (e.g. Analyses).
* PassInstrumentationAnalysis analysis is explicitly requested from PassManager through
usual AnalysisManager::getResult. All pass managers were updated to run that
to get PassInstrumentation object for instrumentation calls.
* Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra
args out of a generic PassManager's extra args. This is the only way I was able to explicitly
run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or
RepeatedPass::run.
TODO: Upon lengthy discussions we agreed to accept this as an initial implementation
and then get rid of getAnalysisResult by improving RepeatedPass implementation.
* PassBuilder takes PassInstrumentationCallbacks object to pass it further into
PassInstrumentationAnalysis. Callbacks registration should be performed directly
through PassInstrumentationCallbacks.
* new-pm tests updated to account for PassInstrumentationAnalysis being run
* Added PassInstrumentation tests to PassBuilderCallbacks unit tests.
Other unit tests updated with registration of the now-required PassInstrumentationAnalysis.
Made getName helper to return std::string (instead of StringRef initially) to fix
asan builtbot failures on CGSCC tests.
Reviewers: chandlerc, philip.pfaffe
Differential Revision: https://reviews.llvm.org/D47858
llvm-svn: 342664
Pass Execution Instrumentation interface enables customizable instrumentation
of pass execution, as per "RFC: Pass Execution Instrumentation interface"
posted 06/07/2018 on llvm-dev@
The intent is to provide a common machinery to implement all
the pass-execution-debugging features like print-before/after,
opt-bisect, time-passes etc.
Here we get a basic implementation consisting of:
* PassInstrumentationCallbacks class that handles registration of callbacks
and access to them.
* PassInstrumentation class that handles instrumentation-point interfaces
that call into PassInstrumentationCallbacks.
* Callbacks accept StringRef which is just a name of the Pass right now.
There were some ideas to pass an opaque wrapper for the pointer to pass instance,
however it appears that pointer does not actually identify the instance
(adaptors and managers might have the same address with the pass they govern).
Hence it was decided to go simple for now and then later decide on what the proper
mental model of identifying a "pass in a phase of pipeline" is.
* Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies
on different IRUnits (e.g. Analyses).
* PassInstrumentationAnalysis analysis is explicitly requested from PassManager through
usual AnalysisManager::getResult. All pass managers were updated to run that
to get PassInstrumentation object for instrumentation calls.
* Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra
args out of a generic PassManager's extra args. This is the only way I was able to explicitly
run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or
RepeatedPass::run.
TODO: Upon lengthy discussions we agreed to accept this as an initial implementation
and then get rid of getAnalysisResult by improving RepeatedPass implementation.
* PassBuilder takes PassInstrumentationCallbacks object to pass it further into
PassInstrumentationAnalysis. Callbacks registration should be performed directly
through PassInstrumentationCallbacks.
* new-pm tests updated to account for PassInstrumentationAnalysis being run
* Added PassInstrumentation tests to PassBuilderCallbacks unit tests.
Other unit tests updated with registration of the now-required PassInstrumentationAnalysis.
Reviewers: chandlerc, philip.pfaffe
Differential Revision: https://reviews.llvm.org/D47858
llvm-svn: 342597
Summary:
Pass Execution Instrumentation interface enables customizable instrumentation
of pass execution, as per "RFC: Pass Execution Instrumentation interface"
posted 06/07/2018 on llvm-dev@
The intent is to provide a common machinery to implement all
the pass-execution-debugging features like print-before/after,
opt-bisect, time-passes etc.
Here we get a basic implementation consisting of:
* PassInstrumentationCallbacks class that handles registration of callbacks
and access to them.
* PassInstrumentation class that handles instrumentation-point interfaces
that call into PassInstrumentationCallbacks.
* Callbacks accept StringRef which is just a name of the Pass right now.
There were some ideas to pass an opaque wrapper for the pointer to pass instance,
however it appears that pointer does not actually identify the instance
(adaptors and managers might have the same address with the pass they govern).
Hence it was decided to go simple for now and then later decide on what the proper
mental model of identifying a "pass in a phase of pipeline" is.
* Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies
on different IRUnits (e.g. Analyses).
* PassInstrumentationAnalysis analysis is explicitly requested from PassManager through
usual AnalysisManager::getResult. All pass managers were updated to run that
to get PassInstrumentation object for instrumentation calls.
* Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra
args out of a generic PassManager's extra args. This is the only way I was able to explicitly
run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or
RepeatedPass::run.
TODO: Upon lengthy discussions we agreed to accept this as an initial implementation
and then get rid of getAnalysisResult by improving RepeatedPass implementation.
* PassBuilder takes PassInstrumentationCallbacks object to pass it further into
PassInstrumentationAnalysis. Callbacks registration should be performed directly
through PassInstrumentationCallbacks.
* new-pm tests updated to account for PassInstrumentationAnalysis being run
* Added PassInstrumentation tests to PassBuilderCallbacks unit tests.
Other unit tests updated with registration of the now-required PassInstrumentationAnalysis.
Reviewers: chandlerc, philip.pfaffe
Differential Revision: https://reviews.llvm.org/D47858
llvm-svn: 342544
invalidation of analyses when merging SCCs.
While I've added a bunch of testing of this, it takes something much
more like the inliner to really trigger this as you need to have
partially-analyzed SCCs with updates at just the right time. So I've
added a direct test for this using the inliner and verifying the
domtree. Without the changes here, this test ends up finding a stale
dominator tree.
However, to handle this properly, we need to invalidate analyses
*before* merging the SCCs. After talking to Philip and Sanjoy about this
they convinced me this was the right approach. To do this, we need
a callback mechanism when merging SCCs so we can observe the cycle that
will be merged before the merge happens. This API update ended up being
surprisingly easy.
With this commit, the new PM passes the test-suite again. It hadn't
since MemorySSA was enabled for EarlyCSE as that also will find this bug
very quickly.
llvm-svn: 307498
the invalidation propagation logic from an SCC to a Function.
I wrote the infrastructure to test this but didn't actually use it in
the unit test where it was designed to be used. =[ My bad. Once
I actually added it to the test case I discovered that it also hadn't
been properly implemented, so I've implemented it. The logic in the FAM
proxy for an SCC pass to propagate invalidation follows the same ideas
as the FAM proxy for a Module pass, but the implementation is a bit
different to reflect the fact that it is forwarding just for an SCC.
However, implementing this correctly uncovered a surprising "bug" (it
was conservatively correct but relatively very expensive) in how we
handle invalidation when splitting one SCC into multiple SCCs. We did an
eager invalidation when in reality we should be deferring invaliadtion
for the *current* SCC to the CGSCC pass manager and just invaliating the
newly constructed SCCs. Otherwise we end up invalidating too much too
soon. This was exposed by the inliner test case that I've updated. Now,
we invalidate *just* the split off '(test1_f)' SCC when doing the CG
update, and then the inliner finishes and invalidates the '(test1_g,
test1_h)' SCC's analyses. The first few attempts at fixing this hit
still more bugs, but all of those are covered by existing tests. For
example, the inliner should also preserve the FAM proxy to avoid
unnecesasry invalidation, and this is safe because the CG update
routines it uses handle any necessary adjustments to the FAM proxy.
Finally, the unittests for the CGSCC pass manager needed a bunch of
updates where we weren't correctly preserving the FAM proxy because it
hadn't been fully implemented and failing to preserve it didn't matter.
Note that this doesn't yet fix the current crasher due to MemSSA finding
a stale dominator tree, but without this the fix to that crasher doesn't
really make any sense when testing because it relies on the proxy
behavior.
llvm-svn: 307487
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
new PM's inliner.
The bug happens when we refine an SCC after having computed a proxy for
the FunctionAnalysisManager, and then proceed to compute fresh analyses
for functions in the *new* SCC using the manager provided by the old
SCC's proxy. *And* when we manage to mutate a function in this new SCC
in a way that invalidates those analyses. This can be... challenging to
reproduce.
I've managed to contrive a set of functions that trigger this and added
a test case, but it is a bit brittle. I've directly checked that the
passes run in the expected ways to help avoid the test just becoming
silently irrelevant.
This gets the new PM back to passing the LLVM test suite after the PGO
improvements landed.
llvm-svn: 292757
Nikita Popov