The metadata added in D102361 introduces a module flag that we can check
to determine if the module was compiled with `-fopenmp` enables. We can
now check for the precense of this instead of scanning the call graph
for OpenMP runtime functions.
Depends on D102361
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102423
Based ontop of D104598, which is a NFCI-ish refactoring.
Here, a restriction, that only empty blocks can be merged, is lifted.
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D104597
After landing the globalization optimizations, the precense of globalization on
the device that was not put in shared or stack memory is a failed optimization
with performance consequences so it should indicate a missed remark.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D104735
Summary:
The changes to globalization introduced in D97680 introduce a large amount of overhead by default. The old globalization method would always ignore globalization code if executing in SPMD mode. This wasn't strictly correct as data sharing is still possible in SPMD mode. The new interface is correct but introduces globalization code even when unnecessary. This optimization will use the existing HeapToStack transformation in the attributor to allow for unneeded globalization to be replaced with thread-private stack memory. This is done using the newly introduced library instances for the RTL functions added in D102087.
Depends on D97818
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D102197
Summary:
Memory globalization is required to maintain OpenMP standard semantics for data sharing between
worker and master threads. The GPU cannot share data between its threads so must allocate global or
shared memory to store the data in. Currently this is implemented fully in the frontend using the
`__kmpc_data_sharing_push_stack` and __kmpc_data_sharing_pop_stack` functions to emulate standard
CPU stack sharing. The front-end scans the target region for variables that escape the region and
must be shared between the threads. Each variable then has a field created for it in a global record
type.
This patch replaces this functinality with a single allocation command, effectively mimicing an
alloca instruction for the variables that must be shared between the threads. This will be much
slower than the current solution, but makes it much easier to optimize as we can analyze each
variable independently and determine if it is not captured. In the future, we can replace these
calls with an `alloca` and small allocations can be pushed to shared memory.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D97680
A backedge-taken count doesn't refer to memory; returning a pointer type
is nonsense. So make sure we always return an integer.
The obvious way to do this would be to just convert the operands of the
icmp to integers, but that doesn't quite work out at the moment:
isLoopEntryGuardedByCond currently gets confused by ptrtoint operations.
So we perform the ptrtoint conversion late for lt/gt operations.
The test changes are mostly innocuous. The most interesting changes are
more complex SCEV expressions of the form "(-1 * (ptrtoint i8* %ptr to
i64)) + %ptr)". This is expected: we can't fold this to zero because we
need to preserve the pointer base.
The call to isLoopEntryGuardedByCond in howFarToZero is less precise
because of ptrtoint operations; this shows up in the function
pr46786_c26_char in ptrtoint.ll. Fixing it here would require more
complex refactoring. It should eventually be fixed by future
improvements to isImpliedCond.
See https://bugs.llvm.org/show_bug.cgi?id=46786 for context.
Differential Revision: https://reviews.llvm.org/D103656
Addition of this pass has been botched.
There is no particular reason why it had to be sold as an inseparable part
of new-pm transition. It was added when old-pm was still the default,
and very *very* few users were actually tracking new-pm,
so it's effects weren't measured.
Which means, some of the turnoil of the new-pm transition
are actually likely regressions due to this pass.
Likewise, there has been a number of post-commit feedback
(post new-pm switch), namely
* https://reviews.llvm.org/D37467#2787157 (regresses HW-loops)
* https://reviews.llvm.org/D37467#2787259 (should not be in middle-end, should run after LSR, not before)
* https://reviews.llvm.org/D95789 (an attempt to fix bad loop backedge metadata)
and in the half year past, the pass authors (google) still haven't found time to respond to any of that.
Hereby it is proposed to backout the pass from the pipeline,
until someone who cares about it can address the issues reported,
and properly start the process of adding a new pass into the pipeline,
with proper performance evaluation.
Furthermore, neither google nor facebook reports any perf changes
from this change, so i'm dropping the pass completely.
It can always be re-reverted should/if anyone want to pick it up again.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D104099
This is split off from D102002, and I think it is clear that
the difference in behavior was not intended. Options were
added to SimplifyCFG over time, but different chunks of
the pass pipelines were not kept in sync.
We can only scalarize memory accesses if we know the index is valid.
This patch adjusts canScalarizeAcceess to fall back to
computeConstantRange to check if the index is known to be valid.
Reviewed By: nlopes
Differential Revision: https://reviews.llvm.org/D102476
This was reverted due to performance regressions in ARM benchmarks,
which have since been addressed by D101196 (SCEV analysis improvement)
and D101778 (CGP reverse transform).
-----
The single-use case is handled implicity by converting the icmp
into a mask check first. When comparing with zero in particular,
we don't need the one-use restriction, as we only produce a single
icmp.
https://alive2.llvm.org/ce/z/MSixcmhttps://alive2.llvm.org/ce/z/GwpG0M
This patch adjusts the LTO pipeline in the new PM to run GlobalsAA
before LICM to match the legacy PM.
This fixes a regression where the new PM failed to vectorize loops that
require hoisting/sinking by LICM depending on GlobalsAA info.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D102345
This is a patch that disables the poison-unsafe select -> and/or i1 folding.
It has been blocking D72396 and also has been the source of a few miscompilations
described in llvm.org/pr49688 .
D99674 conditionally blocked this folding and successfully fixed the latter one.
The former one was still blocked, and this patch addresses it.
Note that a few test functions that has `_logical` suffix are now deoptimized.
These are created by @nikic to check the impact of disabling this optimization
by copying existing original functions and replacing and/or with select.
I can see that most of these are poison-unsafe; they can be revived by introducing
freeze instruction. I left comments at fcmp + select optimizations (or-fcmp.ll, and-fcmp.ll)
because I think they are good targets for freeze fix.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D101191
When passingValueIsAlwaysUndefined scans for an instruction between an
inst with a null or undef argument and its first use, it was checking
for instructions that may have side effects, which is a superset of the
instructions it intended to find (as per the comments, control flow
changing instructions that would prevent reaching the uses). Switch
to using isGuaranteedToTransferExecutionToSuccessor() instead.
Without this change, when enabling -fwhole-program-vtables, which causes
assumes to be inserted by clang, we can get different simplification
decisions. In particular, when building with instrumentation FDO it can
affect the optimizations decisions before FDO matching, leading to some
mismatches.
I had to modify d83507-knowledge-retention-bug.ll since this fix enables
more aggressive optimization of that code such that it no longer tested
the original bug it was meant to test. I removed the undef which still
provokes the original failure (confirmed by temporarily reverting the
fix) and also changed it to just invoke the passes of interest to narrow
the testing.
Similarly I needed to adjust code for UnreachableEliminate.ll to avoid
an undef which was causing the function body to get optimized away with
this fix.
Differential Revision: https://reviews.llvm.org/D101507
Hoisting and sinking instructions out of conditional blocks enables
additional vectorization by:
1. Executing memory accesses unconditionally.
2. Reducing the number of instructions that need predication.
After disabling early hoisting / sinking, we miss out on a few
vectorization opportunities. One of those is causing a ~10% performance
regression in one of the Geekbench benchmarks on AArch64.
This patch tires to recover the regression by running hoisting/sinking
as part of a SimplifyCFG run after LoopRotate and before LoopVectorize.
Note that in the legacy pass-manager, we run LoopRotate just before
vectorization again and there's no SimplifyCFG run in between, so the
sinking/hoisting may impact the later run on LoopRotate. But the impact
should be limited and the benefit of hosting/sinking at this stage
should outweigh the risk of not rotating.
Compile-time impact looks slightly positive for most cases.
http://llvm-compile-time-tracker.com/compare.php?from=2ea7fb7b1c045a7d60fcccf3df3ebb26aa3699e5&to=e58b4a763c691da651f25996aad619cb3d946faf&stat=instructions
NewPM-O3: geomean -0.19%
NewPM-ReleaseThinLTO: geoman -0.54%
NewPM-ReleaseLTO-g: geomean -0.03%
With a few benchmarks seeing a notable increase, but also some
improvements.
Alternative to D101290.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D101468
Summary:
This patch registers OpenMPOpt as a Module pass in addition to a CGSCC
pass. This is so certain optimzations that are sensitive to intact
call-sites can happen before inlining. The old `openmpopt` pass name is
changed to `openmp-opt-cgscc` and `openmp-opt` calls the Module pass.
The current module pass only runs a single check but will be expanded in
the future.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D99202
As a side-effect of the change to default HoistCommonInsts to false
early in the pipeline, we fail to convert conditional branch & phis to
selects early on, which prevents vectorization for loops that contain
conditional branches that effectively are selects (or if the loop gets
vectorized, it will get vectorized very inefficiently).
This patch updates SimplifyCFG to perform hoisting if the only
instruction in both BBs is an equal branch. In this case, the only
additional instructions are selects for phis, which should be cheap.
Even though we perform hoisting, the benefits of this kind of hoisting
should by far outweigh the negatives.
For example, the loop in the code below will not get vectorized on
AArch64 with the current default, but will with the patch. This is a
fundamental pattern we should definitely vectorize. Besides that, I
think the select variants should be easier to use for reasoning across
other passes as well.
https://clang.godbolt.org/z/sbjd8Wshx
```
double clamp(double v) {
if (v < 0.0)
return 0.0;
if (v > 6.0)
return 6.0;
return v;
}
void loop(double* X, double *Y) {
for (unsigned i = 0; i < 20000; i++) {
X[i] = clamp(Y[i]);
}
}
```
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D100329
Retry of 330619a3a6 that includes a clang test update.
Original commit message:
If we run passes before lowering llvm.expect intrinsics to metadata,
then those passes have no way to act on the hints provided by llvm.expect.
SimplifyCFG is the known offender, and we made it smarter about profile
metadata in D98898 <https://reviews.llvm.org/D98898>.
In the motivating example from https://llvm.org/PR49336 , this means we
were ignoring the recommended method for a programmer to tell the compiler
that a compare+branch is expensive. This change appears to solve that case -
the metadata survives to the backend, the compare order is as expected in IR,
and the backend does not do anything to reverse it.
We make the same change to the old pass manager to keep things synchronized.
Differential Revision: https://reviews.llvm.org/D100213
If we run passes before lowering llvm.expect intrinsics to metadata,
then those passes have no way to act on the hints provided by llvm.expect.
SimplifyCFG is the known offender, and we made it smarter about profile
metadata in D98898.
In the motivating example from https://llvm.org/PR49336 , this means we
were ignoring the recommended method for a programmer to tell the compiler
that a compare+branch is expensive. This change appears to solve that case -
the metadata survives to the backend, the compare order is as expected in IR,
and the backend does not do anything to reverse it.
We make the same change to the old pass manager to keep things synchronized.
Differential Revision: https://reviews.llvm.org/D100213
When converting a switch with two cases and a default into a
select, also handle the denegerate case where two cases have the
same value.
Generate this case directly as
%or = or i1 %cmp1, %cmp2
%res = select i1 %or, i32 %val, i32 %default
rather than
%sel1 = select i1 %cmp1, i32 %val, i32 %default
%res = select i1 %cmp2, i32 %val, i32 %sel1
as InstCombine is going to canonicalize to the former anyway.
We handle the case where we have two cases and a default all having
different values, but not the case where two cases happen to have
the same one.
The PhaseOrdering test is a particularly bad example where this
showed up.
This doesn't change anything currently, but as discussed in
D98981 and D98152, some tests may fail to vectorize because
the cost model becomes more accurate as we switch over to
using min/max intrinsics.
When computing a range for a SCEVUnknown, today we use computeKnownBits for unsigned ranges, and computeNumSignBots for signed ranges. This means we miss opportunities to improve range results.
One common missed pattern is that we have a signed range of a value which CKB can determine is positive, but CNSB doesn't convey that information. The current range includes the negative part, and is thus double the size.
Per the removed comment, the original concern which delayed using both (after some code merging years back) was a compile time concern. CTMark results (provided by Nikita, thanks!) showed a geomean impact of about 0.1%. This doesn't seem large enough to avoid higher quality results.
Differential Revision: https://reviews.llvm.org/D96534
Joseph Huber