Summary: visitSwitchInst should not take INT_MAX when Cost is negative. Instead of INT_MAX , we also use a valid upperbound cost when overflow occurs in Cost.
Reviewers: hans, echristo, dmgreen
Reviewed By: dmgreen
Subscribers: mcrosier, javed.absar, llvm-commits, eraman
Differential Revision: https://reviews.llvm.org/D34436
llvm-svn: 306118
Also document the attribute, since "probe-stack" already is.
Reviewed By: majnemer
Differential Revision: https://reviews.llvm.org/D34528
llvm-svn: 306069
This attribute is used to ensure the guard page is triggered on stack
overflow. Stack frames larger than the guard page size will generate
a call to __probestack to touch each page so the guard page won't
be skipped.
Reviewed By: majnemer
Differential Revision: https://reviews.llvm.org/D34386
llvm-svn: 305939
Other comments/implications are that this isn't intended behavior (nor
perserved/reimplemented in the new inliner) & complicates fixing the
'inlining' of trivially dead calls without consulting the cost function
first.
llvm-svn: 305052
Summary:
This is to enable the new switch inline cost heuristic (r301649) by removing the
old heuristic as well as the flag itself.
In my experiment for LLVM test suite and spec2000/2006, +17.82% performance and
8% code size reduce was observed in spec2000/vertex with O3 LTO in AArch64.
No significant code size / performance regression was found in O3/O2/Os. No
significant complain was reported from the llvm-dev thread.
Reviewers: hans, chandlerc, eraman, haicheng, mcrosier, bmakam, eastig, ddibyend, echristo
Reviewed By: echristo
Subscribers: javed.absar, kristof.beyls, echristo, aemerson, rengolin, mehdi_amini
Differential Revision: https://reviews.llvm.org/D32653
llvm-svn: 304594
The added test case is to check whether the simplified value is passed to
getGEPCost().
Differential Revision: https://reviews.llvm.org/D33779
llvm-svn: 304454
Summary:
With instrumentation profiling, when updating the VP metadata after
an inline, VP metadata on the inlined copy was inadvertantly having
all counts zeroed out. This was causing indirect calls from code inlined
during the call step to be marked as cold in the ThinLTO summaries and
not imported.
The CallerBFI needs to be passed down so that the CallSiteCount can be
computed from the profile summary info. With Sample PGO this was working
since the count is extracted from the branch weight metadata on the
call being inlined (even before we stopped looking at metadata for
non-sample PGO in r302844 this largely wasn't working for instrumentation
PGO since only promoted indirect calls would be getting inlined and have
the metadata).
Added an instrumentation PGO test and renamed the sample PGO test.
Reviewers: danielcdh, eraman
Subscribers: mehdi_amini, llvm-commits
Differential Revision: https://reviews.llvm.org/D33389
llvm-svn: 303574
Update threshold based on callee's hotness only when BFI is not available.
Otherwise use only callsite's hotness. This makes it easier to reason about
hotness related threshold updates.
Differential revision: https://reviews.llvm.org/D33157
llvm-svn: 303210
Summary:
Don't use the metadata on call instructions for determining hotness
unless we are in sample PGO mode, where it is needed because profile
counts are not accurate. In instrumentation mode this is not necessary
and does more harm than good when calls have VP metadata that hasn't
been properly scaled after transformations or dropped after constant
prop based devirtualization (both should be fixed, but we don't need
to do this in the first place for instrumentation PGO).
This required adjusting a number of tests to distinguish between sample
and instrumentation PGO handling, and to add in profile summary metadata
so that getProfileCount can get the summary.
Reviewers: davidxl, danielcdh
Subscribers: aemerson, rengolin, mehdi_amini, Prazek, llvm-commits
Differential Revision: https://reviews.llvm.org/D32877
llvm-svn: 302844
I ran the test-suite (including SPEC 2006) in PGO mode comparing cold
thresholds of 225 and 45. Here are some stats on the text size:
Out of 904 tests that ran, 197 see a change in text size. The average
text size reduction (of all the 904 binaries) is 1.07%. Of the 197
binaries, 19 see a text size increase, as high as 18%, but most of them
are small single source benchmarks. There are 3 multisource benchmarks
with a >0.5% size increase (0.7, 1.3 and 2.1 are their % increases). On
the other side of the spectrum, 31 benchmarks see >10% size reduction
and 6 of them are MultiSource.
I haven't run the test-suite with other values of inlinecold-threshold.
Since we have a cold callsite threshold of 45, I picked this value.
Differential revision: https://reviews.llvm.org/D33106
llvm-svn: 302829
Summary:
The motivation example is like below which has 13 cases but only 2 distinct targets
```
lor.lhs.false2: ; preds = %if.then
switch i32 %Status, label %if.then27 [
i32 -7012, label %if.end35
i32 -10008, label %if.end35
i32 -10016, label %if.end35
i32 15000, label %if.end35
i32 14013, label %if.end35
i32 10114, label %if.end35
i32 10107, label %if.end35
i32 10105, label %if.end35
i32 10013, label %if.end35
i32 10011, label %if.end35
i32 7008, label %if.end35
i32 7007, label %if.end35
i32 5002, label %if.end35
]
```
which is compiled into a balanced binary tree like this on AArch64 (similar on X86)
```
.LBB853_9: // %lor.lhs.false2
mov w8, #10012
cmp w19, w8
b.gt .LBB853_14
// BB#10: // %lor.lhs.false2
mov w8, #5001
cmp w19, w8
b.gt .LBB853_18
// BB#11: // %lor.lhs.false2
mov w8, #-10016
cmp w19, w8
b.eq .LBB853_23
// BB#12: // %lor.lhs.false2
mov w8, #-10008
cmp w19, w8
b.eq .LBB853_23
// BB#13: // %lor.lhs.false2
mov w8, #-7012
cmp w19, w8
b.eq .LBB853_23
b .LBB853_3
.LBB853_14: // %lor.lhs.false2
mov w8, #14012
cmp w19, w8
b.gt .LBB853_21
// BB#15: // %lor.lhs.false2
mov w8, #-10105
add w8, w19, w8
cmp w8, #9 // =9
b.hi .LBB853_17
// BB#16: // %lor.lhs.false2
orr w9, wzr, #0x1
lsl w8, w9, w8
mov w9, #517
and w8, w8, w9
cbnz w8, .LBB853_23
.LBB853_17: // %lor.lhs.false2
mov w8, #10013
cmp w19, w8
b.eq .LBB853_23
b .LBB853_3
.LBB853_18: // %lor.lhs.false2
mov w8, #-7007
add w8, w19, w8
cmp w8, #2 // =2
b.lo .LBB853_23
// BB#19: // %lor.lhs.false2
mov w8, #5002
cmp w19, w8
b.eq .LBB853_23
// BB#20: // %lor.lhs.false2
mov w8, #10011
cmp w19, w8
b.eq .LBB853_23
b .LBB853_3
.LBB853_21: // %lor.lhs.false2
mov w8, #14013
cmp w19, w8
b.eq .LBB853_23
// BB#22: // %lor.lhs.false2
mov w8, #15000
cmp w19, w8
b.ne .LBB853_3
```
However, the inline cost model estimates the cost to be linear with the number
of distinct targets and the cost of the above switch is just 2 InstrCosts.
The function containing this switch is then inlined about 900 times.
This change use the general way of switch lowering for the inline heuristic. It
etimate the number of case clusters with the suitability check for a jump table
or bit test. Considering the binary search tree built for the clusters, this
change modifies the model to be linear with the size of the balanced binary
tree. The model is off by default for now :
-inline-generic-switch-cost=false
This change was originally proposed by Haicheng in D29870.
Reviewers: hans, bmakam, chandlerc, eraman, haicheng, mcrosier
Reviewed By: hans
Subscribers: joerg, aemerson, llvm-commits, rengolin
Differential Revision: https://reviews.llvm.org/D31085
llvm-svn: 301649
Summary: Declarations need to be filtered out when counting functions.
Reviewers: eraman
Subscribers: Prazek, llvm-commits
Differential Revision: https://reviews.llvm.org/D31336
llvm-svn: 298720
Summary: Inliner should update the branch_weights annotation to scale it to proper value.
Reviewers: davidxl, eraman
Reviewed By: eraman
Subscribers: zzheng, llvm-commits
Differential Revision: https://reviews.llvm.org/D30767
llvm-svn: 298270
in r297374.
I've extracted a small version of this from the C++ metaprogram Richard
came up with to exercise these kinds of issues and written comments to
describe both how to reproduce a fresh version of the test case and what
likely failure modes are.
The test case is still a bit brittle as it depends on the particular
inline cost modeling and SCC visitation order, but it definitely would
have caught the bug right away when developing things so it seems
a really valuable test case to have.
llvm-svn: 297935
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
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
The LLVM backend cannot produce any debug info for an llvm::Function
without a DISubprogram attachment. When inlining a debug-info-carrying
function into a nodebug function, there is therefore no reason to keep
any debug info intrinsic calls or debug locations on the instructions.
This fixes a problem discovered in PR32042.
rdar://problem/30679307
llvm-svn: 296488
This was suggested in D27855: have the inliner add assumptions, so we don't
lose nonnull info provided by argument attributes.
This still doesn't solve PR28430 (dyn_cast), but this gets us closer.
https://reviews.llvm.org/D29999
llvm-svn: 296366
Multiple blocks in the callee can be mapped to a single cloned block
since we prune the callee as we clone it. The existing code
iterates over the value map and clones the block frequency (and
eventually scales the frequencies of the cloned blocks). Value map's
iteration is not deterministic and so the cloned block might get the
frequency of any of the original blocks. The fix is to set the max of
the original frequencies to the cloned block. The first block in the
sequence must have this max frequency and, in the call context,
subsequent blocks must have its frequency.
Differential Revision: https://reviews.llvm.org/D29696
llvm-svn: 295115
Summary:
As written in the comments above, LastCallToStaticBonus is already applied to
the cost if Caller has only one user, so it is redundant to reapply the bonus
here.
If the only user is not a caller, TotalSecondaryCost will not be adjusted
anyway because callerWillBeRemoved is false. If there's no caller at all, we
don't need to care about TotalSecondaryCost because
inliningPreventsSomeOuterInline is false.
Reviewers: chandlerc, eraman
Reviewed By: eraman
Subscribers: haicheng, davidxl, davide, llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D29169
llvm-svn: 295075
a lazy-asserting PoisoningVH.
AssertVH is fundamentally incompatible with cache-invalidation of
analysis results. The invaliadtion happens after the AssertingVH has
already fired. Instead, use a PoisoningVH that will assert if the
dangling handle is ever used rather than merely be assigned or
destroyed.
This patch also removes all of the (numerous) doomed attempts to work
around this fundamental incompatibility. It is a pretty significant
simplification IMO.
The most interesting change is in the Inliner where we still do some
clearing because we don't want to rely on the coarse grained
invalidation strategy of the containing pass manager. However, I prefer
the approach that contains this logic to the cleanup phase of the
Inliner, and I think we could enhance the CGSCC analysis management
layer to make this even better in the future if desired.
The rest is straight cleanup.
I've also added a test for one of the harder cases to work around: when
a *module analysis* contains many AssertingVHes pointing at functions.
Differential Revision: https://reviews.llvm.org/D29006
llvm-svn: 292928
While this is covered by a clang test case, we should have something
locally to LLVM that immediately checks the inliner doesn't leave
analyses to dangling IR bodies.
llvm-svn: 292772
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
This adds the following to the new PM based inliner in PGO mode:
* Use block frequency analysis to derive callsite's profile count and use
that to adjust thresholds of hot and cold callsites.
* Incrementally update the BFI of the caller after a callee gets inlined
into it. This incremental update is only within an invocation of the run
method - BFI is not preserved across calls to run.
Update the function entry count of the callee after inlining it into a
caller.
* I've tuned the thresholds for the hot and cold callsites using a hacked
up version of the old inliner that explicitly computes BFI on a set of
internal benchmarks and spec. Once the new PM based pipeline stabilizes
(IIRC Chandler mentioned there are known issues) I'll benchmark this
again and adjust the thresholds if required.
Inliner PGO support.
Differential revision: https://reviews.llvm.org/D28331
llvm-svn: 292666
This is the third attemp to recommit r292526.
The original summary:
Currently, a GEP is considered free only if its indices are all constant.
TTI::getGEPCost() can give target-specific more accurate analysis. TTI is
already used for the cost of many other instructions.
llvm-svn: 292633
This is the second attemp to recommit r292526.
The original summary:
Currently, a GEP is considered free only if its indices are all constant.
TTI::getGEPCost() can give target-specific more accurate analysis. TTI is
already used for the cost of many other instructions.
llvm-svn: 292616
This recommits r292526 which is reverted in r292529 after fixing the test case.
The original summary:
Currently, a GEP is considered free only if its indices are all constant.
TTI::getGEPCost() can give target-specific more accurate analysis. TTI is
already used for the cost of many other instructions.
llvm-svn: 292570
Currently, a GEP is considered free only if its indices are all constant.
TTI::getGEPCost() can give target-specific more accurate analysis. TTI is
already used for the cost of many other instructions.
Differential Revision: https://reviews.llvm.org/D28693
llvm-svn: 292526
This is an orthogonal and separated layer instead of being embedded
inside the pass manager. While it adds a small amount of complexity, it
is fairly minimal and the composability and control seems worth the
cost.
The logic for this ends up being nicely isolated and targeted. It should
be easy to experiment with different iteration strategies wrapped around
the CGSCC bottom-up walk using this kind of facility.
The mechanism used to track devirtualization is the simplest one I came
up with. I think it handles most of the cases the existing iteration
machinery handles, but I haven't done a *very* in depth analysis. It
does however match the basic intended semantics, and we can tweak or
tune its exact behavior incrementally as necessary. One thing that we
may want to revisit is freshly building the value handle set on each
iteration. While I don't think this will be a significant cost (it is
strictly fewer value handles but more churn of value handes than the old
call graph), it is conceivable that we'll want a somewhat more clever
tracking mechanism. My hope is to layer that on as a follow up patch
with data supporting any implementation complexity it adds.
This code also provides for a basic count heuristic: if the number of
indirect calls decreases and the number of direct calls increases for
a given function in the SCC, we assume devirtualization is responsible.
This matches the heuristics currently used in the legacy pass manager.
Differential Revision: https://reviews.llvm.org/D23114
llvm-svn: 290665
analyses when we're about to break apart an SCC.
We can't wait until after breaking apart the SCC to invalidate things:
1) Which SCC do we then invalidate? All of them?
2) Even if we invalidate all of them, a newly created SCC may not have
a proxy that will convey the invalidation to functions!
Previously we only invalidated one of the SCCs and too late. This led to
stale analyses remaining in the cache. And because the caching strategy
actually works, they would get used and chaos would ensue.
Doing invalidation early is somewhat pessimizing though if we *know*
that the SCC structure won't change. So it turns out that the design to
make the mutation API force the caller to know the *kind* of mutation in
advance was indeed 100% correct and we didn't do enough of it. So this
change also splits two cases of switching a call edge to a ref edge into
two separate APIs so that callers can clearly test for this and take the
easy path without invalidating when appropriate. This is particularly
important in this case as we expect most inlines to be between functions
in separate SCCs and so the common case is that we don't have to so
aggressively invalidate analyses.
The LCG API change in turn needed some basic cleanups and better testing
in its unittest. No interesting functionality changed there other than
more coverage of the returned sequence of SCCs.
While this seems like an obvious improvement over the current state, I'd
like to revisit the core concept of invalidating within the CG-update
layer at all. I'm wondering if we would be better served forcing the
callers to handle the invalidation beforehand in the cases that they
can handle it. An interesting example is when we want to teach the
inliner to *update and preserve* analyses. But we can cross that bridge
when we get there.
With this patch, the new pass manager an build all of the LLVM test
suite at -O3 and everything passes. =D I haven't bootstrapped yet and
I'm sure there are still plenty of bugs, but this gives a nice baseline
so I'm going to increasingly focus on fleshing out the missing
functionality, especially the bits that are just turned off right now in
order to let us establish this baseline.
llvm-svn: 290664
when they are call edges at the leaf but may (transitively) be reached
via ref edges.
It turns out there is a simple rule: insert everything as a ref edge
which is a safe conservative default. Then we let the existing update
logic handle promoting some of those to call edges.
Note that it would be fairly cheap to make these call edges right away
if that is desirable by testing whether there is some existing call path
from the source to the target. It just seemed like slightly more
complexity in this code path that isn't strictly necessary. If anyone
feels strongly about handling this differently I'm happy to change it.
llvm-svn: 290649
most of the inliner test cases.
The inliner involves a bunch of interesting code and tends to be where
most of the issues I've seen experimenting with the new PM lie. All of
these test cases pass, but I'd like to keep some more thorough coverage
here so doing a fairly blanket enabling.
There are a handful of interesting tests I've not enabled yet because
they're focused on the always inliner, or on functionality that doesn't
(yet) exist in the inliner.
llvm-svn: 290592
skipping indirectly recursive inline chains.
To do this, we implicitly build an inline stack for each callsite and
check prior to inlining that doing so would not form a cycle. This uses
the exact same technique and even shares some code with the legacy PM
inliner.
This solution remains deeply unsatisfying to me because it means we
cannot actually iterate the inliner externally. Doing so would not be
able to easily detect and avoid such cycles. Some day I would very much
like to have a solution that works without this internal state to detect
cycles, but this is not that day.
llvm-svn: 290590
Nothing really interesting here, but I had to improve the test to use
variables rather than hard coding value names as we happen to end up
with different value names in the new PM.
llvm-svn: 290589
Easwaran Raman