As a result of -ftrivial-auto-var-init, clang generates instructions to
set alloca'd memory to a given pattern, right after the allocation site.
In some cases, this (somehow costly) operation could be delayed, leading
to conditional execution in some cases.
This is not an uncommon situation: it happens ~500 times on the cPython
code base, and much more on the LLVM codebase. The benefit greatly
varies on the execution path, but it should not regress on performance.
Differential Revision: https://reviews.llvm.org/D137707
Currently there is no way to enable the module inliner when
linking with full lto. This patch enables that option.
Differential Revision: https://reviews.llvm.org/D146805
This restores commit d6ad4f01c3dafcab335bca66dac6e36d9eac8421, which was
reverted in commit 883dbb9c86be87593a58ef10b070b3a0564c7fee, along with
a fix for gcc 12.2 build errors in the original commit.
Support for building, printing, and displaying CallsiteContextGraph
which represents the MemProf metadata contexts. Uses CRTP to enable
support for both IR (regular LTO) and summary (ThinLTO). This patch
includes the support for building it in regular LTO mode (from
memprof and callsite metadata), and the next patch will add the
handling for building it from ThinLTO summaries.
Also includes support for dumping the graph to text and to dot files.
Follow-on patches will contain the support for cloning on the graph and
in the IR.
The graph represents the call contexts in all memprof metadata on
allocation calls, with nodes for the allocations themselves, as well as
for the calls in each context. The graph is initially built from the
allocation memprof metadata (or summary) MIBs. It is then updated to
match calls with callsite metadata onto the nodes, updating it to
reflect any inlining performed on those calls.
Each MIB (representing an allocation's call context with allocation
behavior) is assigned a unique context id during the graph build. The
edges and nodes in the graph are decorated with the context ids they
carry. This is used to correctly update the graph when cloning is
performed so that we can uniquify the context for a single (possibly
cloned) allocation.
Differential Revision: https://reviews.llvm.org/D140908
Support for building, printing, and displaying CallsiteContextGraph
which represents the MemProf metadata contexts. Uses CRTP to enable
support for both IR (regular LTO) and summary (ThinLTO). This patch
includes the support for building it in regular LTO mode (from
memprof and callsite metadata), and the next patch will add the
handling for building it from ThinLTO summaries.
Also includes support for dumping the graph to text and to dot files.
Follow-on patches will contain the support for cloning on the graph and
in the IR.
The graph represents the call contexts in all memprof metadata on
allocation calls, with nodes for the allocations themselves, as well as
for the calls in each context. The graph is initially built from the
allocation memprof metadata (or summary) MIBs. It is then updated to
match calls with callsite metadata onto the nodes, updating it to
reflect any inlining performed on those calls.
Each MIB (representing an allocation's call context with allocation
behavior) is assigned a unique context id during the graph build. The
edges and nodes in the graph are decorated with the context ids they
carry. This is used to correctly update the graph when cloning is
performed so that we can uniquify the context for a single (possibly
cloned) allocation.
Depends on D140786.
Differential Revision: https://reviews.llvm.org/D140908
The default and pre-link pipeline builders currently require you to
call a separate method for optimization level O0, even though they
have perfectly well-defined O0 optimization pipelines.
Accept O0 optimization level and call buildO0DefaultPipeline()
internally, so all consumers don't need to repeat this.
Differential Revision: https://reviews.llvm.org/D146200
With opaque pointers, all function pointer types are the same, meaning there should be no bitcasts.
Internal benchmarks with SampleFDO look neutral.
This was added in D36333.
Reviewed By: tejohnson, davidxl
Differential Revision: https://reviews.llvm.org/D146099
These are very clearly more simplification than optimization.
Mostly NFC, except for some ordering around passes that don't really matter.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D145967
This is a partial revert of D128830, restoring the previous
position of DeadArgElim in the fat LTO pipeline. The motivation
for this is a major code size regression observed in Rust and
illustrated in the PhaseOrdering test.
This is a conservative fix restoring the previous pipeline order.
The real problem is that the LTO pipeline is conceptually broken:
It doesn't have a CGSCC function simplification pipeline. The
inliner is just being run by itself. This wouldn't be a problem
if fat LTO used a standard design where ArgPromotion and DAE are
only run after functions have already been simplified by the
CGSCC inliner pipeline.
Differential Revision: https://reviews.llvm.org/D146051
This is the follow-up to D144199 and suggestion from D144045.
We make use of loop info explicit via InstCombine pass parameter
rather than semi-arbitrary via caching.
The only InstCombine transform that uses LoopInfo currently is a
GEP fold in visitGEPOfGEP(), so that shows up as a failure in the
dedicated test for the fold as well as several LoopVectorizer tests
that run extra passes.
I don't see any pass manager regression tests that actually check
for pass options, but this is intended to be NFC for the pass
pipeline behavior - we only try to use loop info where it would
have been used before via caching .
Differential Revision: https://reviews.llvm.org/D144274
This is a modified version of commit b374423304a8 by
Arthur (https://reviews.llvm.org/D143424).
Here we invoke to the pass independent of PGOOPT. We now check if the
profile is available through the program summary. This ensures CHR is
called in distributed ThinLTO BE compilation (where PGOOPT might not
be created).
Differential Revision: https://reviews.llvm.org/D144769
This pass isn't a simplification, it's a non-canonical optimization.
This makes it only run once in a (Thin)LTO pipeline during postlink, just like all the other optimization pipeline passes.
Reviewed By: xur
Differential Revision: https://reviews.llvm.org/D143424
This seems to cause large regressions in existing code, as much as 75% slower
(4x the time taken). Small always inline functions seem to be used a lot in the
cmsis-dsp library.
I would add a phase ordering test to show the problems, but one already exists!
The llvm/test/Transforms/PhaseOrdering/ARM/arm_mult_q15.ll was just changed by
removing alwaysinline to hide the problems that existed.
This reverts commit cae033dcf227aeecf58fca5af6fc7fde1fd2fb4f.
This reverts commit 8e33c41e72ad42e4c27f8cbc3ad2e02b169637a1.
We have several situations where it's beneficial for code size to ensure that every
call to always-inline functions are inlined before normal inlining decisions are
made. While the normal inliner runs in a "MandatoryOnly" mode to try to do this,
it only does it on a per-SCC basis, rather than the whole module. Ensuring that
all mandatory inlinings are done before any heuristic based decisions are made
just makes sense.
Despite being referred to the "legacy" AlwaysInliner pass, it's already necessary
for -O0 because the CGSCC inliner is too expensive in compile time to run at -O0.
This also fixes an exponential compile time blow up in
https://github.com/llvm/llvm-project/issues/59126
Differential Revision: https://reviews.llvm.org/D143624
Make the access to profile data going through virtual file system so the
inputs can be remapped. In the context of the caching, it can make sure
we capture the inputs and provided an immutable input as profile data.
Reviewed By: akyrtzi, benlangmuir
Differential Revision: https://reviews.llvm.org/D139052
The `OpenMPOpt` pass is pivotal to the performance of many OpenMP
offloading programs. When we perform non-LTO builds with OpenMP we used
to link the OpenMP deviceRTL individually for each TU. This lead to us
getting an additional attributor run on the combined runtime and user
code. When we used LTO we lost a run and suffered a large performance
degradation. This patch simply adds in the extra `OpenMPOpt` pass that
we miss into the LTO pipeline. This patch fixes the performance
regression shown in applications that used OpenMP offloading in LTO
mode.
Previously, this wasn't legal to do as we could emit new runtime calls
into the module. That was fixed by D142646.
Depends on D142646
Fixes https://github.com/llvm/llvm-project/issues/60300
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D142650
The `OpenMPOpt` pass contains optimizations that generate new calls into
the OpenMP runtime. This causes problems if we are in a state where the
runtime has already been linked statically. Generating these new calls
will result in them never being resolved. We should indicate if we are
in a "post-link" LTO phase and prevent OpenMPOpt from generating new
runtime calls.
Generally, it's not desireable for passes to maintain state about the
context in which they're called. But this is the only reasonable
solution to static linking when we have a pass that generates new
runtime calls.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D142646
This reverts commit fb13dcf3431cd83911fe56899d2fade808dc5b8d.
A large compile-time regression for code generated by sanitizers has
been reported. Revert while I investigate the issue. Details and
reproducers are available here: https://reviews.llvm.org/D135915
Running ConstraintEliminiation after the first InstCombine run results
in slightly more simplifications on average.
There are is a tiny number of regressions, mostly due to CVP eliminating
a condition that ConstraintElimination would use, but in most cases
there's a slight improvement or no change.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D140853
This reverts commit 2656572d485127cc30b8fe9752024d2a0f1c50db.
It looks like CINT2017rate/502.gcc_r gets mis-compiled with LTO + PGO on
AArch64 with function specialization.
This patch enables Function Specialization by default at all
optimization levels except Os, Oz.
Compilation Time Overhead:
--------------------------
Measured the Instruction Count increase (Geomean) for CTMark from
the llvm-testsuite as in https://llvm-compile-time-tracker.com.
* {-O3, Non-LTO}: +0.136% Instruction Count
* {-O3, LTO}: +0.346% Instruction Count
Performance Uplift:
-------------------
Measured +9.121% score increase for 505.mcf_r from SPEC Int 2017
(Tested on Neoverse N1 with -O3 + LTO)
Correctness Testing:
--------------------
* Passes bootstrap Clang with ASAN + LTO + FuncSpec aggressive options:
{ MaxClonesThreshold=10,
SmallFunctionThreshold=10,
AvgLoopIterationCount=30,
SpecializeOnAddresses=true,
EnableSpecializationForLiteralConstant=true,
FuncSpecializationMaxIters=10 }
* Builds Chromium and passes its unittests with the above options + ThinLTO.
For more info please refer to
https://discourse.llvm.org/t/rfc-should-we-enable-function-specialization/61518
Differential Revision: https://reviews.llvm.org/D140210
Reland 877a9f9abec61f06e39f1cd872e37b828139c2d1 since D138654 (parent)
has been fixed with 9ebaf4fef4aac89d4eff08e48185d61bc893f14e and with
8f1e11c5a7d70f96943a72649daa69f152d73e90.
Differential Revision: https://reviews.llvm.org/D126455
Currently, SROA is CFG-preserving.
Not doing so does not affect any pipeline test. (???)
Internally, SROA requires Dominator Tree, and uses it solely for the final `-mem2reg` call.
By design, we can't really SROA alloca if their address escapes somehow,
but we have logic to deal with `load` of `select`/`PHI`,
where at least one of the possible addresses prevents promotion,
by speculating the `load`s and `select`ing between loaded values.
As one would expect, that requires ensuring that the speculation is actually legal.
Even ignoring complexity bailouts, that logic does not deal with everything,
e.g. `isSafeToLoadUnconditionally()` does not recurse into hands of `select`.
There can also be cases where the load is genuinely non-speculate.
So if we can't prove that the load can be speculated,
unfold the select, produce two-entry phi node, and perform predicated load.
Now, that transformation must obviously update Dominator Tree,
since we require it later on. Doing so is trivial.
Additionally, we don't want to do this for the final SROA invocation (D136806).
In the end, this ends up having negative (!) compile-time cost:
https://llvm-compile-time-tracker.com/compare.php?from=c6d7e80ec4c17a415673b1cfd25924f98ac83608&to=ddf9600365093ea50d7e278696cbfa01641c959d&stat=instructions:u
Though indeed, this only deals with `select`s, `PHI`s are still using speculation.
Should we update some more analysis?
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D138238
This reverts commit 739611870d3b06605afe25cc07833f6a62de9545,
and recommits 03e6d9d9d1d48e43f3efc35eb75369b90d4510d5
with a fixed assertion - we should check that DTU is there,
not just assert false...
The assertion about not modifying the CFG seems to not hold,
will recommit in a bit.
https://lab.llvm.org/buildbot#builders/139/builds/32412
This reverts commit 03e6d9d9d1d48e43f3efc35eb75369b90d4510d5.
This reverts commit 4f90f4ada33718f9025d0870a4fe3fe88276b3da.
Currently, SROA is CFG-preserving.
Not doing so does not affect any pipeline test. (???)
Internally, SROA requires Dominator Tree, and uses it solely for the final `-mem2reg` call.
By design, we can't really SROA alloca if their address escapes somehow,
but we have logic to deal with `load` of `select`/`PHI`,
where at least one of the possible addresses prevents promotion,
by speculating the `load`s and `select`ing between loaded values.
As one would expect, that requires ensuring that the speculation is actually legal.
Even ignoring complexity bailouts, that logic does not deal with everything,
e.g. `isSafeToLoadUnconditionally()` does not recurse into hands of `select`.
There can also be cases where the load is genuinely non-speculate.
So if we can't prove that the load can be speculated,
unfold the select, produce two-entry phi node, and perform predicated load.
Now, that transformation must obviously update Dominator Tree,
since we require it later on. Doing so is trivial.
Additionally, we don't want to do this for the final SROA invocation (D136806).
In the end, this ends up having negative (!) compile-time cost:
https://llvm-compile-time-tracker.com/compare.php?from=c6d7e80ec4c17a415673b1cfd25924f98ac83608&to=ddf9600365093ea50d7e278696cbfa01641c959d&stat=instructions:u
Though indeed, this only deals with `select`s, `PHI`s are still using speculation.
Should we update some more analysis?
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D138238
This reverts commit 877a9f9abec61f06e39f1cd872e37b828139c2d1.
It depends on the parent revision 42c2dc401742266da3e0251b6c1ca491f4779963
which needs to be reverted as it broke some buildbots, so reverting both.
The aim of this patch is to minimize the compilation time overhead of
running Function Specialization. It is about 40% slower to run as a
standalone pass (IPSCCP + FuncSpec vs IPSCCP with FuncSpec) according
to my measurements. I compiled the llvm testsuite with NewPM-O3 + LTO
and measured single threaded [user + system] time of IPSCCP and FuncSpec
by passing the '-time-passes' option to lld. Then I compared the two
configurations in terms of Instruction Count of the total compilation
(not of the individual passes) as in https://llvm-compile-time-tracker.com.
Geomean for non-LTO builds is -0.25% and LTO is -0.5% approximately.
You can find more info below:
https://discourse.llvm.org/t/rfc-should-we-enable-function-specialization/61518
Differential Revision: https://reviews.llvm.org/D126455
ControlHeightReduction (CHR) clones the code region to reduce the
branches in the hot code path. The number of clones is linear to the
depth of the region.
Currently it does not have control over the code size increase. We are
seeing one ~9000 BB functions get expanded to ~250000 BBs, an 25x
increase. This creates a big compile time issue for the downstream
optimizations.
This patch adds a cap for number of clones for one region.
Differential Revision: https://reviews.llvm.org/D138333
The option was added with https://reviews.llvm.org/D102496,
and currently the name is accurate, but I am hoping to add
a load transform that is not a scalarization. See issue #17113.
Move these to the new PM if they're used there.
Part of removing the legacy pass manager for optimization pipeline.
Reland with UseNewGVN usage in clang removed.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D137915
Move these to the new PM if they're used there.
Part of removing the legacy pass manager for optimization pipeline.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D137915
