The primitive conservative heuristic seems to give a slight overall
improvement while not regressing stuff. Make it available to wider
testing. If you notice any speed regressions (or significant code
size regressions) let me know!
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156258 91177308-0d34-0410-b5e6-96231b3b80d8
This came up when a change in block placement formed a cmov and slowed down a
hot loop by 50%:
ucomisd (%rdi), %xmm0
cmovbel %edx, %esi
cmov is a really bad choice in this context because it doesn't get branch
prediction. If we emit it as a branch, an out-of-order CPU can do a better job
(if the branch is predicted right) and avoid waiting for the slow load+compare
instruction to finish. Of course it won't help if the branch is unpredictable,
but those are really rare in practice.
This patch uses a dumb conservative heuristic, it turns all cmovs that have one
use and a direct memory operand into branches. cmovs usually save some code
size, so we disable the transform in -Os mode. In-Order architectures are
unlikely to benefit as well, those are included in the
"predictableSelectIsExpensive" flag.
It would be better to reuse branch probability info here, but BPI doesn't
support select instructions currently. It would make sense to use the same
heuristics as the if-converter pass, which does the opposite direction of this
transform.
Test suite shows a small improvement here and there on corei7-level machines,
but the actual results depend a lot on the used microarchitecture. The
transformation is currently disabled by default and available by passing the
-enable-cgp-select2branch flag to the code generator.
Thanks to Chandler for the initial test case to him and Evan Cheng for providing
me with comments and test-suite numbers that were more stable than mine :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156234 91177308-0d34-0410-b5e6-96231b3b80d8
of the CodeExtractor utility. This allows speculatively computing input
and output sets to measure the likely size impact of the code
extraction.
These sets cannot be reused sadly -- we mutate the function prior to
forming the final sets used by the actual extraction.
The interface has been revamped slightly to make it easier to use
correctly by making the interface const and sinking the computation of
the number of exit blocks into the full extraction function and away
from the rest of this logic which just computed two output parameters.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156168 91177308-0d34-0410-b5e6-96231b3b80d8
blocks, assert that this doesn't happen. We don't want to bother trying
to support this call pattern as it isn't necessary.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156167 91177308-0d34-0410-b5e6-96231b3b80d8
and expose it as a utility class rather than as free function wrappers.
The simple free-function interface works well for the bugpoint-specific
pass's uses of code extraction, but in an upcoming patch for more
advanced code extraction, they simply don't expose a rich enough
interface. I need to expose various stages of the process of doing the
code extraction and query information to decide whether or not to
actually complete the extraction or give up.
Rather than build up a new predicate model and pass that into these
functions, just take the class that was actually implementing the
functions and lift it up into a proper interface that can be used to
perform code extraction. The interface is cleaned up and re-documented
to work better in a header. It also is now setup to accept the blocks to
be extracted in the constructor rather than in a method.
In passing this essentially reverts my previous commit here exposing
a block-level query for eligibility of extraction. That is no longer
necessary with the more rich interface as clients can query the
extraction object for eligibility directly. This will reduce the number
of walks of the input basic block sequence by quite a bit which is
useful if this enters the normal optimization pipeline.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156163 91177308-0d34-0410-b5e6-96231b3b80d8
minor behavior changes with this, but nothing I have seen evidence of in
the wild or expect to be meaningful. The real goal is unifying our logic
and simplifying the interfaces. A summary of the changes follows:
- Make 'callIsSmall' actually accept a callsite so it can handle
intrinsics, and simplify callers appropriately.
- Nuke a completely bogus declaration of 'callIsSmall' that was still
lurking in InlineCost.h... No idea how this got missed.
- Teach the 'isInstructionFree' about the various more intelligent
'free' heuristics that got added to the inline cost analysis during
review and testing. This mostly surrounds int->ptr and ptr->int casts.
- Switch most of the interesting parts of the inline cost analysis that
were essentially computing 'is this instruction free?' to use the code
metrics routine instead. This way we won't keep duplicating logic.
All of this is motivated by the desire to allow other passes to compute
a roughly equivalent 'cost' metric for a particular basic block as the
inline cost analysis. Sadly, re-using the same analysis for both is
really messy because only the actual inline cost analysis is ever going
to go to the contortions required for simplification, SROA analysis,
etc.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156140 91177308-0d34-0410-b5e6-96231b3b80d8
extraction into a public interface. Also clean it up and apply it more
consistently such that we check for landing pads *anywhere* in the
extracted code, not just in single-block extraction.
This will be used to guide decisions in passes that are planning to
eventually perform a round of code extraction.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156114 91177308-0d34-0410-b5e6-96231b3b80d8
<rdar://problem/11291436>.
This is a second attempt at a fix for this, the first was r155468. Thanks
to Chandler, Bob and others for the feedback that helped me improve this.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155866 91177308-0d34-0410-b5e6-96231b3b80d8
Allow the "SplitCriticalEdge" function to split the edge to a landing pad. If
the pass is *sure* that it thinks it knows what it's doing, then it may go ahead
and specify that the landing pad can have its critical edge split. The loop
unswitch pass is one of these passes. It will split the critical edges of all
edges coming from a loop to a landing pad not within the loop. Doing so will
retain important loop analysis information, such as loop simplify.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155817 91177308-0d34-0410-b5e6-96231b3b80d8
Target specific types should not be vectorized. As a practical matter,
these types are already register matched (at least in the x86 case),
and codegen does not always work correctly (at least in the ppc case,
and this is not worth fixing because ppc_fp128 is currently broken and
will probably go away soon).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155729 91177308-0d34-0410-b5e6-96231b3b80d8
The limit is set to an arbitrary 1000 recursion depth to avoid stack overflow
issues. <rdar://problem/11286839>.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155722 91177308-0d34-0410-b5e6-96231b3b80d8
The required checks are moved to ChainInstruction() itself and the
policy decisions are moved to IVChain::isProfitableInc().
Also cache the ExprBase in IVChain to avoid frequent recomputations.
No functional change intended.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155676 91177308-0d34-0410-b5e6-96231b3b80d8
elements to minimize the number of multiplies required to compute the
final result. This uses a heuristic to attempt to form near-optimal
binary exponentiation-style multiply chains. While there are some cases
it misses, it seems to at least a decent job on a very diverse range of
inputs.
Initial benchmarks show no interesting regressions, and an 8%
improvement on SPASS. Let me know if any other interesting results (in
either direction) crop up!
Credit to Richard Smith for the core algorithm, and helping code the
patch itself.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155616 91177308-0d34-0410-b5e6-96231b3b80d8
of a precise count. Also, move RRInfo's Partial field into PtrState,
now that it won't increase the size.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155513 91177308-0d34-0410-b5e6-96231b3b80d8
These lists exclude invoke unwind edges and loop backedges which
are being ignored. This makes it easier to ignore them
consistently.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155500 91177308-0d34-0410-b5e6-96231b3b80d8
Original commit message:
Defer some shl transforms to DAGCombine.
The shl instruction is used to represent multiplication by a constant
power of two as well as bitwise left shifts. Some InstCombine
transformations would turn an shl instruction into a bit mask operation,
making it difficult for later analysis passes to recognize the
constsnt multiplication.
Disable those shl transformations, deferring them to DAGCombine time.
An 'shl X, C' instruction is now treated mostly the same was as 'mul X, C'.
These transformations are deferred:
(X >>? C) << C --> X & (-1 << C) (When X >> C has multiple uses)
(X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2) (When C2 > C1)
(X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2) (When C1 > C2)
The corresponding exact transformations are preserved, just like
div-exact + mul:
(X >>?,exact C) << C --> X
(X >>?,exact C1) << C2 --> X << (C2-C1)
(X >>?,exact C1) << C2 --> X >>?,exact (C1-C2)
The disabled transformations could also prevent the instruction selector
from recognizing rotate patterns in hash functions and cryptographic
primitives. I have a test case for that, but it is too fragile.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155362 91177308-0d34-0410-b5e6-96231b3b80d8
While the patch was perfect and defect free, it exposed a really nasty
bug in X86 SelectionDAG that caused an llc crash when compiling lencod.
I'll put the patch back in after fixing the SelectionDAG problem.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155181 91177308-0d34-0410-b5e6-96231b3b80d8
The shl instruction is used to represent multiplication by a constant
power of two as well as bitwise left shifts. Some InstCombine
transformations would turn an shl instruction into a bit mask operation,
making it difficult for later analysis passes to recognize the
constsnt multiplication.
Disable those shl transformations, deferring them to DAGCombine time.
An 'shl X, C' instruction is now treated mostly the same was as 'mul X, C'.
These transformations are deferred:
(X >>? C) << C --> X & (-1 << C) (When X >> C has multiple uses)
(X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2) (When C2 > C1)
(X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2) (When C1 > C2)
The corresponding exact transformations are preserved, just like
div-exact + mul:
(X >>?,exact C) << C --> X
(X >>?,exact C1) << C2 --> X << (C2-C1)
(X >>?,exact C1) << C2 --> X >>?,exact (C1-C2)
The disabled transformations could also prevent the instruction selector
from recognizing rotate patterns in hash functions and cryptographic
primitives. I have a test case for that, but it is too fragile.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155136 91177308-0d34-0410-b5e6-96231b3b80d8
If the loop contains invoke instructions, whose unwind edge escapes the loop,
then don't try to unswitch the loop. Doing so may cause the unwind edge to be
split, which not only is non-trivial but doesn't preserve loop simplify
information.
Fixes PR12573
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154987 91177308-0d34-0410-b5e6-96231b3b80d8
This introduces a threshold of 200 IV Users, which is very
conservative but should be sufficient to avoid serious compile time
sink or stack overflow. The llvm test-suite with LTO never exceeds 190
users per loop.
The bug doesn't relate to a specific type of loop. Checking in an
arbitrary giant loop as a unit test would be silly.
Fixes rdar://11262507.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154983 91177308-0d34-0410-b5e6-96231b3b80d8
also fix SimplifyLibCalls to use TLI rather than compile-time conditionals to enable optimizations on floor, ceil, round, rint, and nearbyint
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154960 91177308-0d34-0410-b5e6-96231b3b80d8
When vectorizing pointer types it is important to realize that potential
pairs cannot be connected via the address pointer argument of a load or store.
This is because even after vectorization, the address is still a scalar because
the address of the higher half of the pair is implicit from the address of the
lower half (it need not be, and should not be, explicitly computed).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154735 91177308-0d34-0410-b5e6-96231b3b80d8
As has been suggested by Duncan and others, Early-CSE and GVN should
do similar redundancy elimination, but Early-CSE is much less expensive.
Most of my autovectorization benchmarks show a performance regresion, but
all of these are < 0.1%, and so I think that it is still worth using
the less expensive pass.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154673 91177308-0d34-0410-b5e6-96231b3b80d8
library return value optimization for phi uses. Even when the
phi itself is not dominated, the specific use may be dominated.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154647 91177308-0d34-0410-b5e6-96231b3b80d8
obviously cannot know that this code is present, let alone used. So prevent the
internalize pass from internalizing those global values which code-gen may
insert.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154645 91177308-0d34-0410-b5e6-96231b3b80d8
- don't isntrument reads from constant globals.
Saves ~1.5% of instrumented instructions on CPU2006
(counting static instructions, not their execution).
- don't insrument reads from vtable (which is a global constant too).
Saves ~5%.
I did not measure the run-time impact of this,
but it is certainly non-negative.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154444 91177308-0d34-0410-b5e6-96231b3b80d8
a write to the same temp follows in the same BB.
Also add stats printing.
On Spec CPU2006 this optimization saves roughly 4% of instrumented reads
(which is 3% of all instrumented accesses):
Writes : 161216
Reads : 446458
Reads-before-write: 18295
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154418 91177308-0d34-0410-b5e6-96231b3b80d8
Take this opportunity to generalize the indirectbr bailout logic for
loop transformations. CFG transformations will never get indirectbr
right, and there's no point trying.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154386 91177308-0d34-0410-b5e6-96231b3b80d8
GEPs, bit casts, and stores reaching it but no other instructions. These
often show up during the iterative processing of the inliner, SROA, and
DCE. Once we hit this point, we can completely remove the alloca. These
were actually showing up in the final, fully optimized code in a bunch
of inliner tests I've been working on, and notably they show up after
LLVM finishes optimizing away all function calls involved in
hash_combine(a, b).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154285 91177308-0d34-0410-b5e6-96231b3b80d8
simplification has been performed. This is a bit less efficient
(requires another ilist walk of the basic blocks) but shouldn't matter
in practice. More importantly, it's just too much work to keep track of
all the various ways the return instructions can be mutated while
simplifying them. This fixes yet another crasher, reported by Daniel
Dunbar.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154179 91177308-0d34-0410-b5e6-96231b3b80d8
dead code, including dead return instructions in some cases. Otherwise,
we end up having a bogus poniter to a return instruction that blows up
much further down the road.
It turns out that this pattern is both simpler to code, easier to update
in the face of enhancements to the inliner cleanup, and likely cheaper
given that it won't add dead instructions to the list.
Thanks to John Regehr's numerous test cases for teasing this out.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154157 91177308-0d34-0410-b5e6-96231b3b80d8
of the BBVectorizePass without using command line option. As pointed out
by Hal, we can ask the TargetLoweringInfo for the architecture specific
VectorizeConfig to perform vectorizing with architecture specific
information.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154096 91177308-0d34-0410-b5e6-96231b3b80d8
BasicBlock in other passes, e.g. we can call vectorizeBasicBlock in the
loop unroll pass right after the loop is unrolled.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154089 91177308-0d34-0410-b5e6-96231b3b80d8
LSR can fold three addressing modes into its ICmpZero node:
ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset
ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset
ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
The first two cases are only used if TLI->isLegalICmpImmediate() likes
the offset.
Make sure the right Offset sign is passed to this method in the second
case. The ARM version is not symmetric.
<rdar://problem/11184260>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154079 91177308-0d34-0410-b5e6-96231b3b80d8
This allows us to keep passing reduced masks to SimplifyDemandedBits, but
know about all the bits if SimplifyDemandedBits fails. This allows instcombine
to simplify cases like the one in the included testcase.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154011 91177308-0d34-0410-b5e6-96231b3b80d8
reducing unroll count, otherwise the reduced unroll count is not taking
the "OptimizeForSize" attribute into account.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@154007 91177308-0d34-0410-b5e6-96231b3b80d8
http://llvm.org/bugs/show_bug.cgi?id=12343
We have not trivial way for splitting edges that are goes from indirect branch. We can do it with some tricks, but it should be additionally discussed. And it is still dangerous due to difficulty of indirect branches controlling.
Fix forbids this case for unswitching.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153879 91177308-0d34-0410-b5e6-96231b3b80d8
As a side note, I really dislike array_pod_sort... Do we really still
care about any STL implementations that get this so wrong? Does libc++?
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153834 91177308-0d34-0410-b5e6-96231b3b80d8
a single missing character. Somehow, this had gone untested. I've added
tests for returns-twice logic specifically with the always-inliner that
would have caught this, and fixed the bug.
Thanks to Matt for the careful review and spotting this!!! =D
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153832 91177308-0d34-0410-b5e6-96231b3b80d8
the very high overhead of the complex inline cost analysis when all it
wants to do is detect three patterns which must not be inlined. Comment
the code, clean it up, and leave some hints about possible performance
improvements if this ever shows up on a profile.
Moving this off of the (now more expensive) inline cost analysis is
particularly important because we have to run this inliner even at -O0.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153814 91177308-0d34-0410-b5e6-96231b3b80d8
interfaces. These methods were used in the old inline cost system where
there was a persistent cache that had to be updated, invalidated, and
cleared. We're now doing more direct computations that don't require
this intricate dance. Even if we resume some level of caching, it would
almost certainly have a simpler and more narrow interface than this.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153813 91177308-0d34-0410-b5e6-96231b3b80d8
on a per-callsite walk of the called function's instructions, in
breadth-first order over the potentially reachable set of basic blocks.
This is a major shift in how inline cost analysis works to improve the
accuracy and rationality of inlining decisions. A brief outline of the
algorithm this moves to:
- Build a simplification mapping based on the callsite arguments to the
function arguments.
- Push the entry block onto a worklist of potentially-live basic blocks.
- Pop the first block off of the *front* of the worklist (for
breadth-first ordering) and walk its instructions using a custom
InstVisitor.
- For each instruction's operands, re-map them based on the
simplification mappings available for the given callsite.
- Compute any simplification possible of the instruction after
re-mapping, and store that back int othe simplification mapping.
- Compute any bonuses, costs, or other impacts of the instruction on the
cost metric.
- When the terminator is reached, replace any conditional value in the
terminator with any simplifications from the mapping we have, and add
any successors which are not proven to be dead from these
simplifications to the worklist.
- Pop the next block off of the front of the worklist, and repeat.
- As soon as the cost of inlining exceeds the threshold for the
callsite, stop analyzing the function in order to bound cost.
The primary goal of this algorithm is to perfectly handle dead code
paths. We do not want any code in trivially dead code paths to impact
inlining decisions. The previous metric was *extremely* flawed here, and
would always subtract the average cost of two successors of
a conditional branch when it was proven to become an unconditional
branch at the callsite. There was no handling of wildly different costs
between the two successors, which would cause inlining when the path
actually taken was too large, and no inlining when the path actually
taken was trivially simple. There was also no handling of the code
*path*, only the immediate successors. These problems vanish completely
now. See the added regression tests for the shiny new features -- we
skip recursive function calls, SROA-killing instructions, and high cost
complex CFG structures when dead at the callsite being analyzed.
Switching to this algorithm required refactoring the inline cost
interface to accept the actual threshold rather than simply returning
a single cost. The resulting interface is pretty bad, and I'm planning
to do lots of interface cleanup after this patch.
Several other refactorings fell out of this, but I've tried to minimize
them for this patch. =/ There is still more cleanup that can be done
here. Please point out anything that you see in review.
I've worked really hard to try to mirror at least the spirit of all of
the previous heuristics in the new model. It's not clear that they are
all correct any more, but I wanted to minimize the change in this single
patch, it's already a bit ridiculous. One heuristic that is *not* yet
mirrored is to allow inlining of functions with a dynamic alloca *if*
the caller has a dynamic alloca. I will add this back, but I think the
most reasonable way requires changes to the inliner itself rather than
just the cost metric, and so I've deferred this for a subsequent patch.
The test case is XFAIL-ed until then.
As mentioned in the review mail, this seems to make Clang run about 1%
to 2% faster in -O0, but makes its binary size grow by just under 4%.
I've looked into the 4% growth, and it can be fixed, but requires
changes to other parts of the inliner.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153812 91177308-0d34-0410-b5e6-96231b3b80d8
The powi intrinsic requires special handling because it always takes a single
integer power regardless of the result type. As a result, we can vectorize
only if the powers are equal. Fixes PR12364.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153797 91177308-0d34-0410-b5e6-96231b3b80d8
CodeGenPrepare sinks compare instructions down to their uses to prevent
live flags and predicate registers across basic blocks.
PRE of a compare instruction prevents that, forcing the i1 compare
result into a general purpose register. That is usually more expensive
than the redundant compare PRE was trying to eliminate in the first
place.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153657 91177308-0d34-0410-b5e6-96231b3b80d8
blocks in the function cloner. This removes the last case of trivially
dead code that I've been seeing in the wild getting inlined, analyzed,
re-inlined, optimized, only to be deleted. Nukes a FIXME from the
cleanup tests.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153572 91177308-0d34-0410-b5e6-96231b3b80d8
size bloat. Unfortunately, I expect this to disable the majority of the
benefit from r152737. I'm hopeful at least that it will fix PR12345. To
explain this requires... quite a bit of backstory I'm afraid.
TL;DR: The change in r152737 actually did The Wrong Thing for
linkonce-odr functions. This change makes it do the right thing. The
benefits we saw were simple luck, not any actual strategy. Benchmark
numbers after a mini-blog-post so that I've written down my thoughts on
why all of this works and doesn't work...
To understand what's going on here, you have to understand how the
"bottom-up" inliner actually works. There are two fundamental modes to
the inliner:
1) Standard fixed-cost bottom-up inlining. This is the mode we usually
think about. It walks from the bottom of the CFG up to the top,
looking at callsites, taking information about the callsite and the
called function and computing th expected cost of inlining into that
callsite. If the cost is under a fixed threshold, it inlines. It's
a touch more complicated than that due to all the bonuses, weights,
etc. Inlining the last callsite to an internal function gets higher
weighth, etc. But essentially, this is the mode of operation.
2) Deferred bottom-up inlining (a term I just made up). This is the
interesting mode for this patch an r152737. Initially, this works
just like mode #1, but once we have the cost of inlining into the
callsite, we don't just compare it with a fixed threshold. First, we
check something else. Let's give some names to the entities at this
point, or we'll end up hopelessly confused. We're considering
inlining a function 'A' into its callsite within a function 'B'. We
want to check whether 'B' has any callers, and whether it might be
inlined into those callers. If so, we also check whether inlining 'A'
into 'B' would block any of the opportunities for inlining 'B' into
its callers. We take the sum of the costs of inlining 'B' into its
callers where that inlining would be blocked by inlining 'A' into
'B', and if that cost is less than the cost of inlining 'A' into 'B',
then we skip inlining 'A' into 'B'.
Now, in order for #2 to make sense, we have to have some confidence that
we will actually have the opportunity to inline 'B' into its callers
when cheaper, *and* that we'll be able to revisit the decision and
inline 'A' into 'B' if that ever becomes the correct tradeoff. This
often isn't true for external functions -- we can see very few of their
callers, and we won't be able to re-consider inlining 'A' into 'B' if
'B' is external when we finally see more callers of 'B'. There are two
cases where we believe this to be true for C/C++ code: functions local
to a translation unit, and functions with an inline definition in every
translation unit which uses them. These are represented as internal
linkage and linkonce-odr (resp.) in LLVM. I enabled this logic for
linkonce-odr in r152737.
Unfortunately, when I did that, I also introduced a subtle bug. There
was an implicit assumption that the last caller of the function within
the TU was the last caller of the function in the program. We want to
bonus the last caller of the function in the program by a huge amount
for inlining because inlining that callsite has very little cost.
Unfortunately, the last caller in the TU of a linkonce-odr function is
*not* the last caller in the program, and so we don't want to apply this
bonus. If we do, we can apply it to one callsite *per-TU*. Because of
the way deferred inlining works, when it sees this bonus applied to one
callsite in the TU for 'B', it decides that inlining 'B' is of the
*utmost* importance just so we can get that final bonus. It then
proceeds to essentially force deferred inlining regardless of the actual
cost tradeoff.
The result? PR12345: code bloat, code bloat, code bloat. Another result
is getting *damn* lucky on a few benchmarks, and the over-inlining
exposing critically important optimizations. I would very much like
a list of benchmarks that regress after this change goes in, with
bitcode before and after. This will help me greatly understand what
opportunities the current cost analysis is missing.
Initial benchmark numbers look very good. WebKit files that exhibited
the worst of PR12345 went from growing to shrinking compared to Clang
with r152737 reverted.
- Bootstrapped Clang is 3% smaller with this change.
- Bootstrapped Clang -O0 over a single-source-file of lib/Lex is 4%
faster with this change.
Please let me know about any other performance impact you see. Thanks to
Nico for reporting and urging me to actually fix, Richard Smith, Duncan
Sands, Manuel Klimek, and Benjamin Kramer for talking through the issues
today.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@153506 91177308-0d34-0410-b5e6-96231b3b80d8
