The earlier usage of wouldInstructionBeTriviallyDead is based on the
assumption that the use_count of that instruction being checked will be
zero. This patch separates the API into two different ones:
1. The strictly conservative one where the instruction is trivially dead iff the uses are dead.
2. The slightly relaxed form, where an instruction is dead along paths where it is not used.
The second form can be used in identifying instructions that are valid
to sink down to uses (D109917).
Reviewed-By: reames
Differential Revision: https://reviews.llvm.org/D114647
This is a continuation of D109860 and D109903.
An important challenge for profile inference is caused by the fact that the
sample profile is collected on a fully optimized binary, while the block and
edge frequencies are consumed on an early stage of the compilation that operates
with a non-optimized IR. As a result, some of the basic blocks may not have
associated sample counts, and it is up to the algorithm to deduce missing
frequencies. The problem is illustrated in the figure where three basic
blocks are not present in the optimized binary and hence, receive no samples
during profiling.
We found that it is beneficial to treat all such blocks equally. Otherwise the
compiler may decide that some blocks are “cold” and apply undesirable
optimizations (e.g., hot-cold splitting) regressing the performance. Therefore,
we want to distribute the counts evenly along the blocks with missing samples.
This is achieved by a post-processing step that identifies "dangling" subgraphs
consisting of basic blocks with no sampled counts; once the subgraphs are
found, we rebalance the flow so as every branch probability is 50:50 within the
subgraphs.
Our experiments indicate up to 1% performance win using the optimization on
some binaries and a significant improvement in the quality of profile counts
(when compared to ground-truth instrumentation-based counts)
{F19093045}
Reviewed By: hoy
Differential Revision: https://reviews.llvm.org/D109980
This is a continuation of D109860.
Traditional flow-based algorithms cannot guarantee that the resulting edge
frequencies correspond to a *connected* flow in the control-flow graph. For
example, for an instance in the attached figure, a flow-based (or any other)
inference algorithm may produce an output in which the hot loop is disconnected
from the entry block (refer to the rightmost graph in the figure). Furthermore,
creating a connected minimum-cost maximum flow is a computationally NP-hard
problem. Hence, we apply a post-processing adjustments to the computed flow
by connecting all isolated flow components ("islands").
This feature helps to keep all blocks with sample counts connected and results
in significant performance wins for some binaries.
{F19077343}
Reviewed By: hoy
Differential Revision: https://reviews.llvm.org/D109903
When doing load/store promotion within LICM, if we
cannot prove that it is safe to sink the store we won't
hoist the load, even though we can prove the load could
be dereferenced and moved outside the loop. This patch
implements the load promotion by moving it in the loop
preheader by inserting proper PHI in the loop. The store
is kept as is in the loop. By doing this, we avoid doing
the load from a memory location in each iteration.
Please consider this small example:
loop {
var = *ptr;
if (var) break;
*ptr= var + 1;
}
After this patch, it will be:
var0 = *ptr;
loop {
var1 = phi (var0, var2);
if (var1) break;
var2 = var1 + 1;
*ptr = var2;
}
This addresses some problems from [0].
[0] https://bugs.llvm.org/show_bug.cgi?id=51193
Differential revision: https://reviews.llvm.org/D113289
Add support for memset_pattern{4,8} similar to the existing
memset_pattern16 handling.
Reviewed By: ab
Differential Revision: https://reviews.llvm.org/D114883
This fixes the assertion failure reported in https://reviews.llvm.org/D114889#3166417,
by making RecursivelyDeleteTriviallyDeadInstructionsPermissive()
more permissive. As the function accepts a WeakTrackingVH, even if
originally only Instructions were inserted, we may end up with
different Value types after a RAUW operation. As such, we should
not assume that the vector only contains instructions.
Notably this matches the behavior of the
RecursivelyDeleteTriviallyDeadInstructions() function variant which
accepts a single value rather than vector.
`memcpy_chk` can be treated like `memcpy`, with the exception that it
may not return (if it aborts the program).
See D114793 for a similar patch for `memset_chk`.
Reviewed By: xbolva00
Differential Revision: https://reviews.llvm.org/D114863
Previously we missed cloning metadata on function declarations because
we don't call CloneFunctionInto() on declarations in CloneModule().
Reviewed By: dexonsmith
Differential Revision: https://reviews.llvm.org/D113812
The benefits of sampling-based PGO crucially depends on the quality of profile
data. This diff implements a flow-based algorithm, called profi, that helps to
overcome the inaccuracies in a profile after it is collected.
Profi is an extended and significantly re-engineered classic MCMF (min-cost
max-flow) approach suggested by Levin, Newman, and Haber [2008, Complementing
missing and inaccurate profiling using a minimum cost circulation algorithm]. It
models profile inference as an optimization problem on a control-flow graph with
the objectives and constraints capturing the desired properties of profile data.
Three important challenges that are being solved by profi:
- "fixing" errors in profiles caused by sampling;
- converting basic block counts to edge frequencies (branch probabilities);
- dealing with "dangling" blocks having no samples in the profile.
The main implementation (and required docs) are in SampleProfileInference.cpp.
The worst-time complexity is quadratic in the number of blocks in a function,
O(|V|^2). However a careful engineering and extensive evaluation shows that
the running time is (slightly) super-linear. In particular, instances with
1000 blocks are solved within 0.1 second.
The algorithm has been extensively tested internally on prod workloads,
significantly improving the quality of generated profile data and providing
speedups in the range from 0% to 5%. For "smaller" benchmarks (SPEC06/17), it
generally improves the performance (with a few outliers) but extra work in
the compiler might be needed to re-tune existing optimization passes relying on
profile counts.
UPD Dec 1st 2021:
- synced the declaration and definition of the option `SampleProfileUseProfi ` to use type `cl::opt<bool`;
- added `inline` for `SampleProfileInference<BT>::findUnlikelyJumps` and `SampleProfileInference<BT>::isExit` to avoid linking problems on windows.
Reviewed By: wenlei, hoy
Differential Revision: https://reviews.llvm.org/D109860
When doing load/store promotion within LICM, if we
cannot prove that it is safe to sink the store we won't
hoist the load, even though we can prove the load could
be dereferenced and moved outside the loop. This patch
implements the load promotion by moving it in the loop
preheader by inserting proper PHI in the loop. The store
is kept as is in the loop. By doing this, we avoid doing
the load from a memory location in each iteration.
Please consider this small example:
loop {
var = *ptr;
if (var) break;
*ptr= var + 1;
}
After this patch, it will be:
var0 = *ptr;
loop {
var1 = phi (var0, var2);
if (var1) break;
var2 = var1 + 1;
*ptr = var2;
}
This addresses some problems from [0].
[0] https://bugs.llvm.org/show_bug.cgi?id=51193
Differential revision: https://reviews.llvm.org/D113289
The memset_chk library function should match memset's attributes with
respect of memory effects (argmemonly, writeonly). It also does not
raise exceptions. It may not return, in case it aborts the program.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D114793
The basic problem we have is that we're trying to reuse an instruction which is mapped to some SCEV. Since we can have multiple such instructions (potentially with different flags), this is analogous to our need to drop flags when performing CSE. A trivial implementation would simply drop flags on any instruction we decided to reuse, and that would be correct.
This patch is almost that trivial patch except that we preserve flags on the reused instruction when existing users would imply UB on overflow already. Adding new users can, at most, refine this program to one which doesn't execute UB which is valid.
In practice, this fixes two conceptual problems with the previous code: 1) a binop could have been canonicalized into a form with different opcode or operands, or 2) the inbounds GEP case which was simply unhandled.
On the test changes, most are pretty straight forward. We loose some flags (in some cases, they'd have been dropped on the next CSE pass anyways). The one that took me the longest to understand was the ashr-expansion test. What's happening there is that we're considering reuse of the mul, previously we disallowed it entirely, now we allow it with no flags. The surrounding diffs are all effects of generating the same mul with a different operand order, and then doing simple DCE.
The loss of the inbounds is unfortunate, but even there, we can recover most of those once we actually treat branch-on-poison as immediate UB.
Differential Revision: https://reviews.llvm.org/D112734
This solves a problem with non-deterministic output from opt due
to not performing dominator tree updates in a deterministic order.
The problem that was analysed indicated that JumpThreading was using
the DomTreeUpdater via llvm::MergeBasicBlockIntoOnlyPred. When
preparing the list of updates to send to DomTreeUpdater::applyUpdates
we iterated over a SmallPtrSet, which didn't give a well-defined
order of updates to perform.
The added domtree-updates.ll test case is an example that would
result in non-deterministic printouts of the domtree. Semantically
those domtree:s are equivalent, but it show the fact that when we
use the domtree iterator the order in which nodes are visited depend
on the order in which dominator tree updates are performed.
Since some passes (at least EarlyCSE) are iterating over nodes in the
dominator tree in a similar fashion as the domtree printer, then the
order in which transforms are applied by such passes, transitively,
also depend on the order in which dominator tree updates are
performed. And taking EarlyCSE as an example the end result could be
different depending on in which order the transforms are applied.
Reviewed By: nikic, kuhar
Differential Revision: https://reviews.llvm.org/D110292
The benefits of sampling-based PGO crucially depends on the quality of profile
data. This diff implements a flow-based algorithm, called profi, that helps to
overcome the inaccuracies in a profile after it is collected.
Profi is an extended and significantly re-engineered classic MCMF (min-cost
max-flow) approach suggested by Levin, Newman, and Haber [2008, Complementing
missing and inaccurate profiling using a minimum cost circulation algorithm]. It
models profile inference as an optimization problem on a control-flow graph with
the objectives and constraints capturing the desired properties of profile data.
Three important challenges that are being solved by profi:
- "fixing" errors in profiles caused by sampling;
- converting basic block counts to edge frequencies (branch probabilities);
- dealing with "dangling" blocks having no samples in the profile.
The main implementation (and required docs) are in SampleProfileInference.cpp.
The worst-time complexity is quadratic in the number of blocks in a function,
O(|V|^2). However a careful engineering and extensive evaluation shows that
the running time is (slightly) super-linear. In particular, instances with
1000 blocks are solved within 0.1 second.
The algorithm has been extensively tested internally on prod workloads,
significantly improving the quality of generated profile data and providing
speedups in the range from 0% to 5%. For "smaller" benchmarks (SPEC06/17), it
generally improves the performance (with a few outliers) but extra work in
the compiler might be needed to re-tune existing optimization passes relying on
profile counts.
Reviewed By: wenlei, hoy
Differential Revision: https://reviews.llvm.org/D109860
The benefits of sampling-based PGO crucially depends on the quality of profile
data. This diff implements a flow-based algorithm, called profi, that helps to
overcome the inaccuracies in a profile after it is collected.
Profi is an extended and significantly re-engineered classic MCMF (min-cost
max-flow) approach suggested by Levin, Newman, and Haber [2008, Complementing
missing and inaccurate profiling using a minimum cost circulation algorithm]. It
models profile inference as an optimization problem on a control-flow graph with
the objectives and constraints capturing the desired properties of profile data.
Three important challenges that are being solved by profi:
- "fixing" errors in profiles caused by sampling;
- converting basic block counts to edge frequencies (branch probabilities);
- dealing with "dangling" blocks having no samples in the profile.
The main implementation (and required docs) are in SampleProfileInference.cpp.
The worst-time complexity is quadratic in the number of blocks in a function,
O(|V|^2). However a careful engineering and extensive evaluation shows that
the running time is (slightly) super-linear. In particular, instances with
1000 blocks are solved within 0.1 second.
The algorithm has been extensively tested internally on prod workloads,
significantly improving the quality of generated profile data and providing
speedups in the range from 0% to 5%. For "smaller" benchmarks (SPEC06/17), it
generally improves the performance (with a few outliers) but extra work in
the compiler might be needed to re-tune existing optimization passes relying on
profile counts.
Reviewed By: wenlei, hoy
Differential Revision: https://reviews.llvm.org/D109860
The if-check above deleted part guarantees that StoreOffset <= LoadOffset
and that StoreOffset + StoreSize >= LoadOffset + LoadSize, and given that
LoadOffset + LoadSize > LoadOffset when LoadSize > 0. Thus, this shows
StoreOffset + StoreSize > LoadOffset is guaranteed given LoadSize > 0,
while it could be meaningless to have a type with nonpositive size, so that
the check could be removed. The values are converted to signed types to
avoid unsigned operation with negative offsets.
Part of revision D100179
Reapply commit c35e8185d8 with fixing problem
reported by mstorsjo
There are still another 2 uses of PointerType::getElementType in
Evaluator when evaluating BitCast's on pointers. BitCast's on pointers
should be removed when opaque ptr is ready, so I just keep them as is.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D114131
This change is mostly about getting rid of some "uninteresting" cases in a follow on deeper heuristic change. If anyone sees actually interesting code differences out of this, please let me know. I'm not expecting this to have much impact at all.
Case 1 - With the single deoptimize non-latch exit, we can't have two exiting blocks sharing an exit block. We can only hit this with a poorly documented debug flag.
Case 2 - Why should we treat epilog cases differently from prolog cases? Or to say it differently, why should starting with a constant control whether a multiple exit loop gets unrolled?
Sorry for the lack of tests here. These are both *exceedingly* narrow cases in practice, and after a while trying, I couldn't come up with a test which did anything "useful" as opposed to simply exercise a random combination of force flags. Note that the legality cases for each are already exercised with force flags.
All of the interesting logic from this routine has been removed, inline the single check into the sole non-assert caller. The assert use has little value with the restructured code and is simply dropped.
This reverts commit c35e8185d8.
mstorsjo reported in the revision thread that one VNCoercion assertion
is violated and seemly in relate to this commit. As per "If a test case
that demonstrates a problem is reported in the commit thread, please
revert and investigate offline", this commit is reverted.
ProfileCount could model invalid values, but a user had no indication
that the getCount method could return bogus data. Optional<ProfileCount>
addresses that, because the user must dereference the optional. In
addition, the patch removes concept duplication.
Differential Revision: https://reviews.llvm.org/D113839
The if-check above deleted part guarantees that StoreOffset <= LoadOffset
and that StoreOffset + StoreSize >= LoadOffset + LoadSize, and given that
LoadOffset + LoadSize > LoadOffset when LoadSize > 0. Thus, this shows
StoreOffset + StoreSize > LoadOffset is guaranteed given LoadSize > 0,
while it could be meaningless to have a type with nonpositive size, so that
the check could be removed.
Part of revision D100179
Reviewed By: nikic
This is one of those wonderful "in theory X doesn't matter, but in practice is does" changes. In this particular case, we shift the IVs inserted by the runtime unroller to clamp iteration count of the loops* from decrementing to incrementing.
Why does this matter? A couple of reasons:
* SCEV doesn't have a native subtract node. Instead, all subtracts (A - B) are represented as A + -1 * B and drops any flags invalidated by such. As a result, SCEV is slightly less good at reasoning about edge cases involving decrementing addrecs than incrementing ones. (You can see this in the inferred flags in some of the test cases.)
* Other parts of the optimizer produce incrementing IVs, and they're common in idiomatic source language. We do have support for reversing IVs, but in general if we produce one of each, the pair will persist surprisingly far through the optimizer before being coalesced. (You can see this looking at nearby phis in the test cases.)
Note that if the hardware prefers decrementing (i.e. zero tested) loops, LSR should convert back immediately before codegen.
* Mostly irrelevant detail: The main loop of the prolog case is handled independently and will simple use the original IV with a changed start value. We could in theory use this scheme for all iteration clamping, but that's a larger and more invasive change.
The unrolling code was previously inserting new cloned blocks at the end of the function. The result of this with typical loop structures is that the new iterations are placed far from the initial iteration.
With unrolling, the general assumption is that the a) the loop is reasonable hot, and b) the first Count-1 copies of the loop are rarely (if ever) loop exiting. As such, placing Count-1 copies out of line is a fairly poor code placement choice. We'd much rather fall through into the hot (non-exiting) path. For code with branch profiles, later layout would fix this, but this may have a positive impact on non-PGO compiled code.
However, the real motivation for this change isn't performance. Its readability and human understanding. Having to jump around long distances in an IR file to trace an unrolled loop structure is error prone and tedious.
Without this patch, passingValueIsAlwaysUndefined will iterate over all
instructions from I to the end of the basic block, even if the use is
outside the block.
This patch adds an early bail out, if the use instruction is outside I's
BB. This can greatly reduce compile-time in cases where very large basic
blocks are involved, with a large number of PHI nodes and incoming
values.
Note that the refactoring makes the handling of the case where I is a
phi and Use is in PHI more explicit as well: for phi nodes, we can also
directly bail out. In the existing code, we would iterate until we reach
the end and return false.
Based on an earlier patch by Matt Wala.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D113293
This is a fix for test failures on expensive checks build caused by db289340c8.
With LLVM_ENABLE_EXPENSIVE_CHECKS enabled the llvm::sort shuffles the given container.
However, the sort is only called when the TTI is passed to replaceCongruentIVs.
In the mentioned patch we pass it TTI, so the sort happens. But due to shuffling
equivalent Phis may appear in different order from run to run.
With the stable_sort instead of sort this is impossible - the order of sorted Phis
is preserved.
Extended value is known to be inside range smaller than full one.
Prevent SCCP to mark such value as overdefined.
Fixes PR52253
Differential Revision: https://reviews.llvm.org/D112721
Added support for peeling loops with exits that are followed either by an
unreachable-terminated block or block that has a terminatnig deoptimize call.
All blocks in the sequence must have an unique successor, maybe except
for the last one.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D110922
It's a no-op, no overflow happens ever: https://alive2.llvm.org/ce/z/Zw89rZ
While generally i don't like such hacks,
we have a very good reason to do this: here we are expanding
a run-time correctness check for the vectorization,
and said `umul_with_overflow` will not be optimized out
before we query the cost of the checks we've generated.
Which means, the cost of run-time checks would be artificially inflated,
and after https://reviews.llvm.org/D109368 that will affect
the minimal trip count for which these checks are even evaluated.
And if they aren't even evaluated, then the vectorized code
certainly won't be run.
We could consider doing this in IRBuilder, but then we'd need to
also teach `CreateExtractValue()` to look into chain of `insertvalue`'s,
and i'm not sure there's precedent for that.
Refs. https://reviews.llvm.org/D109368#3089809
The function simplifyOnce only calls simplifyOnceImpl and does nothing else.
Having this separate helper makes no sense. Removing it.
Patch by Dmitry Bakunevich!
Differential Revision: https://reviews.llvm.org/D112517
Reviewed By: mkazantsev
Anna Thomas