Nearly all the changes to this pass have been done while maintaining and
updating other parts of LLVM. LLVM has had another pass, SROA, which
has superseded ScalarReplAggregates for quite some time.
Differential Revision: http://reviews.llvm.org/D21316
llvm-svn: 272737
Summary:
The module pass pipeline includes a late LICM run after loop
unrolling. LCSSA is implicitly run as a pass dependency of LICM. However no
cleanup pass was run after this, so the LCSSA nodes ended in the optimized output.
Reviewers: hfinkel, mehdi_amini
Subscribers: majnemer, bruno, mzolotukhin, mehdi_amini, llvm-commits
Differential Revision: http://reviews.llvm.org/D20606
llvm-svn: 271602
As a result of D18634 we no longer infer certain attributes on linkonce_odr
functions at compile time, and may only infer them at LTO time. The readnone
attribute in particular is required for virtual constant propagation (part
of whole-program virtual call optimization) to work correctly.
This change moves the whole-program virtual call optimization pass after
the function attribute inference passes, and enables the attribute inference
passes at opt level 1, so that virtual constant propagation has a chance to
work correctly for linkonce_odr functions.
Differential Revision: http://reviews.llvm.org/D20643
llvm-svn: 270765
Summary:
The original ThinLTO pipeline was derived from some
work I did tuning FullLTO on the test suite and SPEC. This
patch reduces the amount of work done in the "linker phase" of
the build, and extend the function simplifications passes
performed during the "compile phase". This helps the build time
by reducing the IR as much as possible during the compile phase
and limiting the work to be performed during the "link phase",
while keeping the performance "on par" with the existing pipeline.
Reviewers: tejohnson
Subscribers: llvm-commits, joker.eph
Differential Revision: http://reviews.llvm.org/D19773
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 268769
This pass is supposed to reduce the size of the IR for compile time
purpose. We should run it ASAP, except when we prepare for LTO or
ThinLTO, and we want to keep them available for link-time inline.
Differential Revision: http://reviews.llvm.org/D19813
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 268394
This is where it was originally, until LoopVersioningLICM was
inserted before in r259986, I don't believe it was on purpose.
Differential Revision: http://reviews.llvm.org/D19809
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 268252
This patch implements the transformation that promotes indirect calls to
conditional direct calls when the indirect-call value profile meta-data is
available.
Differential Revision: http://reviews.llvm.org/D17864
llvm-svn: 267815
Summary:
D19403 adds a new pragma for loop distribution. This change adds
support for the corresponding metadata that the pragma is translated to
by the FE.
As part of this I had to rethink the flag -enable-loop-distribute. My
goal was to be backward compatible with the existing behavior:
A1. pass is off by default from the optimization pipeline
unless -enable-loop-distribute is specified
A2. pass is on when invoked directly from opt (e.g. for unit-testing)
The new pragma/metadata overrides these defaults so the new behavior is:
B1. A1 + enable distribution for individual loop with the pragma/metadata
B2. A2 + disable distribution for individual loop with the pragma/metadata
The default value whether the pass is on or off comes from the initiator
of the pass. From the PassManagerBuilder the default is off, from opt
it's on.
I moved -enable-loop-distribute under the pass. If the flag is
specified it overrides the default from above.
Then the pragma/metadata can further modifies this per loop.
As a side-effect, we can now also use -enable-loop-distribute=0 from opt
to emulate the default from the optimization pipeline. So to be precise
this is the new behavior:
C1. pass is off by default from the optimization pipeline
unless -enable-loop-distribute or the pragma/metadata enables it
C2. pass is on when invoked directly from opt
unless -enable-loop-distribute=0 or the pragma/metadata disables it
Reviewers: hfinkel
Subscribers: joker.eph, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D19431
llvm-svn: 267672
Summary:
This IR pass is helpful for GPUs, and other targets with divergent
branches. It's a nop on targets without divergent branches.
Reviewers: chandlerc
Subscribers: llvm-commits, jingyue, rnk, joker.eph, tra
Differential Revision: http://reviews.llvm.org/D18626
llvm-svn: 266399
Summary:
For correct handling of alias to nameless
function, we need to be able to refer them through a GUID in the summary.
Here we name them using a hash of the non-private global names in the module.
Reviewers: tejohnson
Subscribers: joker.eph, llvm-commits
Differential Revision: http://reviews.llvm.org/D18883
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 266132
Summary:
This gives callers flexibility to pass lambdas with captures, which lets
callers avoid the C-style void*-ptr closure style. (Currently, callers
in clang store state in the PassManagerBuilderBase arg.)
No functional change, and the new API is backwards-compatible.
Reviewers: chandlerc
Subscribers: joker.eph, cfe-commits
Differential Revision: http://reviews.llvm.org/D18613
llvm-svn: 264918
The latent bug that LLE exposed in the LoopVectorizer was resolved
(PR26952).
The pass can be disabled with -mllvm -enable-loop-load-elim=0
llvm-svn: 263595
(Resubmitting after fixing missing file issue)
With the changes in r263275, there are now more than just functions in
the summary. Completed the renaming of data structures (started in
r263275) to reflect the wider scope. In particular, changed the
FunctionIndex* data structures to ModuleIndex*, and renamed related
variables and comments. Also renamed the files to reflect the changes.
A companion clang patch will immediately succeed this patch to reflect
this renaming.
llvm-svn: 263513
With the changes in r263275, there are now more than just functions in
the summary. Completed the renaming of data structures (started in
r263275) to reflect the wider scope. In particular, changed the
FunctionIndex* data structures to ModuleIndex*, and renamed related
variables and comments. Also renamed the files to reflect the changes.
A companion clang patch will immediately succeed this patch to reflect
this renaming.
llvm-svn: 263490
tests to run GVN in both modes.
This is mostly the boring refactoring just like SROA and other complex
transformation passes. There is some trickiness in that GVN's
ValueNumber class requires hand holding to get to compile cleanly. I'm
open to suggestions about a better pattern there, but I tried several
before settling on this. I was trying to balance my desire to sink as
much implementation detail into the source file as possible without
introducing overly many layers of abstraction.
Much like with SROA, the design of this system is made somewhat more
cumbersome by the need to support both pass managers without duplicating
the significant state and logic of the pass. The same compromise is
struck here.
I've also left a FIXME in a doxygen comment as the GVN pass seems to
have pretty woeful documentation within it. I'd like to submit this with
the FIXME and let those more deeply familiar backfill the information
here now that we have a nice place in an interface to put that kind of
documentaiton.
Differential Revision: http://reviews.llvm.org/D18019
llvm-svn: 263208
As part of r251146 InstCombine was extended to call computeKnownBits on
every value in the function to determine whether it happens to be
constant. This increases typical compiletime by 1-3% (5% in irgen+opt
time) in my measurements. On the other hand this case did not trigger
once in the whole llvm-testsuite.
This patch introduces the notion of ExpensiveCombines which are only
enabled for OptLevel > 2. I removed the check in InstructionSimplify as
that is called from various places where the OptLevel is not known but
given the rarity of the situation I think a check in InstCombine is
enough.
Differential Revision: http://reviews.llvm.org/D16835
llvm-svn: 263047
This reverts commit r262250.
It causes SPEC2006/gcc to generate wrong result (166.s) in AArch64 when
running with *ref* data set. The error happens with
"-Ofast -flto -fuse-ld=gold" or "-O3 -fno-strict-aliasing".
llvm-svn: 262839
Summary:
I re-benchmarked this and results are similar to original results in
D13259:
On ARM64:
SingleSource/Benchmarks/Polybench/linear-algebra/solvers/dynprog -59.27%
SingleSource/Benchmarks/Polybench/stencils/adi -19.78%
On x86:
SingleSource/Benchmarks/Polybench/linear-algebra/solvers/dynprog -27.14%
And of course the original ~20% gain on SPECint_2006/456.hmmer with Loop
Distribution.
In terms of compile time, there is ~5% increase on both
SingleSource/Benchmarks/Misc/oourafft and
SingleSource/Benchmarks/Linkpack/linkpack-pc. These are both very tiny
loop-intensive programs where SCEV computations dominates compile time.
The reason that time spent in SCEV increases has to do with the design
of the old pass manager. If a transform pass does not preserve an
analysis we *invalidate* the analysis even if there was *no*
modification made by the transform pass.
This means that currently we don't take advantage of LLE and LV sharing
the same analysis (LAA) and unfortunately we recompute LAA *and* SCEV
for LLE.
(There should be a way to work around this limitation in the case of
SCEV and LAA since both compute things on demand and internally cache
their result. Thus we could pretend that transform passes preserve
these analyses and manually invalidate them upon actual modification.
On the other hand the new pass manager is supposed to solve so I am not
sure if this is worthwhile.)
Reviewers: hfinkel, dberlin
Subscribers: dberlin, reames, mssimpso, aemerson, joker.eph, llvm-commits
Differential Revision: http://reviews.llvm.org/D16300
llvm-svn: 262250
convert one test to use this.
This is a particularly significant milestone because it required
a working per-function AA framework which can be queried over each
function from within a CGSCC transform pass (and additionally a module
analysis to be accessible). This is essentially *the* point of the
entire pass manager rewrite. A CGSCC transform is able to query for
multiple different function's analysis results. It works. The whole
thing appears to actually work and accomplish the original goal. While
we were able to hack function attrs and basic-aa to "work" in the old
pass manager, this port doesn't use any of that, it directly leverages
the new fundamental functionality.
For this to work, the CGSCC framework also has to support SCC-based
behavior analysis, etc. The only part of the CGSCC pass infrastructure
not sorted out at this point are the updates in the face of inlining and
running function passes that mutate the call graph.
The changes are pretty boring and boiler-plate. Most of the work was
factored into more focused preperatory patches. But this is what wires
it all together.
llvm-svn: 261203
Summary:
On the contrary to Full LTO, ThinLTO can afford to shift compile time
from the frontend to the linker: both phases are parallel (even if
it is not totally "free": projects like clang are reusing product
from the "compile phase" for multiple link, think about
libLLVMSupport reused for opt, llc, etc.).
This pipeline is based on the proposal in D13443 for full LTO. We
didn't move forward on this proposal because the LTO link was far too
long after that. We believe that we can afford it with ThinLTO.
The ThinLTO pipeline integrates in the regular O2/O3 flow:
- The compile phase perform the inliner with a somehow lighter
function simplification. (TODO: tune the inliner thresholds here)
This is intendend to simplify the IR and get rid of obvious things
like linkonce_odr that will be inlined.
- The link phase will run the pipeline from the start, extended with
some specific passes that leverage the augmented knowledge we have
during LTO. Especially after the inliner is done, a sequence of
globalDCE/globalOpt is performed, followed by another run of the
"function simplification" passes. It is not clear if this part
of the pipeline will stay as is, as the split model of ThinLTO
does not allow the same benefit as FullLTO without added tricks.
The measurements on the public test suite as well as on our internal
suite show an overall net improvement. The binary size for the clang
executable is reduced by 5%. We're still tuning it with the bringup
of ThinLTO and it will evolve, but this should provide a good starting
point.
Reviewers: tejohnson
Differential Revision: http://reviews.llvm.org/D17115
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 261029
It is intended to contains the passes run over a function after the
inliner is done with a function and before it moves to its callers.
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 261028
Summary:
On the contrary to Full LTO, ThinLTO can afford to shift compile time
from the frontend to the linker: both phases are parallel.
This pipeline is based on the proposal in D13443 for full LTO. We ]
didn't move forward on this proposal because the link was far too long
after that.
This patch refactor the "function simplification" passes that are part
of the inliner loop in a helper function (this part is NFC and can be
commited separately to simplify the diff). The ThinLTO pipeline
integrates in the regular O2/O3 flow:
- The compile phase perform the inliner with a somehow lighter
function simplification. (TODO: tune the inliner thresholds here)
This is intendend to simplify the IR and get rid of obvious things
like linkonce_odr that will be inlined.
- The link phase will run the pipeline from the start, extended with
some specific passes that leverage the augmented knowledge we have
during LTO. Especially after the inliner is done, a sequence of
globalDCE/globalOpt is performed, followed by another run of the
"function simplification" passes.
The measurements on the public test suite as well as on our internal
suite show an overall net improvement. The binary size for the clang
executable is reduced by 5%. We're still tuning it with the bringup
of ThinLTO but this should provide a good starting point.
Reviewers: tejohnson
Subscribers: joker.eph, llvm-commits, dexonsmith
Differential Revision: http://reviews.llvm.org/D17115
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 260604
It is intended to contains the passes run over a function after the
inliner is done with a function and before it moves to its callers.
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 260603
This pass implements whole program optimization of virtual calls in cases
where we know (via bitset information) that the list of callees is fixed. This
includes the following:
- Single implementation devirtualization: if a virtual call has a single
possible callee, replace all calls with a direct call to that callee.
- Virtual constant propagation: if the virtual function's return type is an
integer <=64 bits and all possible callees are readnone, for each class and
each list of constant arguments: evaluate the function, store the return
value alongside the virtual table, and rewrite each virtual call as a load
from the virtual table.
- Uniform return value optimization: if the conditions for virtual constant
propagation hold and each function returns the same constant value, replace
each virtual call with that constant.
- Unique return value optimization for i1 return values: if the conditions
for virtual constant propagation hold and a single vtable's function
returns 0, or a single vtable's function returns 1, replace each virtual
call with a comparison of the vptr against that vtable's address.
Differential Revision: http://reviews.llvm.org/D16795
llvm-svn: 260312
Summary:
When alias analysis is uncertain about the aliasing between any two accesses,
it will return MayAlias. This uncertainty from alias analysis restricts LICM
from proceeding further. In cases where alias analysis is uncertain we might
use loop versioning as an alternative.
Loop Versioning will create a version of the loop with aggressive aliasing
assumptions in addition to the original with conservative (default) aliasing
assumptions. The version of the loop making aggressive aliasing assumptions
will have all the memory accesses marked as no-alias. These two versions of
loop will be preceded by a memory runtime check. This runtime check consists
of bound checks for all unique memory accessed in loop, and it ensures the
lack of memory aliasing. The result of the runtime check determines which of
the loop versions is executed: If the runtime check detects any memory
aliasing, then the original loop is executed. Otherwise, the version with
aggressive aliasing assumptions is used.
The pass is off by default and can be enabled with command line option
-enable-loop-versioning-licm.
Reviewers: hfinkel, anemet, chatur01, reames
Subscribers: MatzeB, grosser, joker.eph, sanjoy, javed.absar, sbaranga,
llvm-commits
Differential Revision: http://reviews.llvm.org/D9151
llvm-svn: 259986
This patch includes the passmanagerbuilder change that enables IR level PGO instrumentation. It adds two passmanagerbuilder options: -profile-generate=<profile_filename> and -profile-use=<profile_filename>. The new options are primarily for debug purpose.
Reviewers: davidxl, silvas
Differential Revision: http://reviews.llvm.org/D15828
llvm-svn: 258420
Loop trip counts can often be resolved during LTO. We should obviously be unrolling small loops once those trip counts have been resolved, but we weren't.
llvm-svn: 257767
a top-down manner into a true top-down or RPO pass over the call graph.
There are specific patterns of function attributes, notably the
norecurse attribute, which are most effectively propagated top-down
because all they us caller information.
Walk in RPO over the call graph SCCs takes the form of a module pass run
immediately after the CGSCC pass managers postorder walk of the SCCs,
trying again to deduce norerucrse for each singular SCC in the call
graph.
This removes a very legacy pass manager specific trick of using a lazy
revisit list traversed during finalization of the CGSCC pass. There is
no analogous finalization step in the new pass manager, and a lazy
revisit list is just trying to produce an RPO iteration of the call
graph. We can do that more directly if more expensively. It seems
unlikely that this will be the expensive part of any compilation though
as we never examine the function bodies here. Even in an LTO run over
a very large module, this should be a reasonable fast set of operations
over a reasonably small working set -- the function call graph itself.
In the future, if this really is a compile time performance issue, we
can look at building support for both post order and RPO traversals
directly into a pass manager that builds and maintains the PO list of
SCCs.
Differential Revision: http://reviews.llvm.org/D15785
llvm-svn: 257163
a standalone pass.
There is no call graph or even interesting analysis for this part of
function attributes -- it is literally inferring attributes based on the
target library identification. As such, we can do it using a much
simpler module pass that just walks the declarations. This can also
happen much earlier in the pass pipeline which has benefits for any
number of other passes.
In the process, I've cleaned up one particular aspect of the logic which
was necessary in order to separate the two passes cleanly. It now counts
inferred attributes independently rather than just counting all the
inferred attributes as one, and the counts are more clearly explained.
The two test cases we had for this code path are both ... woefully
inadequate and copies of each other. I've kept the superset test and
updated it. We need more testing here, but I had to pick somewhere to
stop fixing everything broken I saw here.
Differential Revision: http://reviews.llvm.org/D15676
llvm-svn: 256466
is (by default) run much earlier than FuncitonAttrs proper.
This allows forcing optnone or other widely impactful attributes. It is
also a bit simpler as the force attribute behavior needs no specific
iteration order.
I've added the pass into the default module pass pipeline and LTO pass
pipeline which mirrors where function attrs itself was being run.
Differential Revision: http://reviews.llvm.org/D15668
llvm-svn: 256465
An LTO pass that generates a __cfi_check() function that validates a
call based on a hash of the call-site-known type and the target
pointer.
llvm-svn: 255693
This patch does two things:
1. mem2reg is now run immediately after globalopt. Now that globalopt
can localize variables more aggressively, it makes sense to lower
them to SSA form earlier rather than later so they can benefit from
the full set of optimization passes.
2. More scalar optimizations are run after the loop optimizations in
LTO mode. The loop optimizations (especially indvars) can clean up
scalar code sufficiently to make it worthwhile running more scalar
passes. I've particularly added SCCP here as it isn't run anywhere
else in the LTO pass pipeline.
Mem2reg is super cheap and shouldn't affect compilation time at all. The
rest of the added passes are in the LTO pipeline only so doesn't affect
the vast majority of compilations, just the link step.
llvm-svn: 255634
Summary:
Add a field on the PassManagerBuilder that clang or gold can use to pass
down a pointer to the function index in memory to use for importing when
the ThinLTO backend is triggered. Add support to supply this to the
function import pass.
Reviewers: joker.eph, dexonsmith
Subscribers: davidxl, llvm-commits, joker.eph
Differential Revision: http://reviews.llvm.org/D15024
llvm-svn: 254926
Summary:
The goal of this pass is to perform store-to-load forwarding across the
backedge of a loop. E.g.:
for (i)
A[i + 1] = A[i] + B[i]
=>
T = A[0]
for (i)
T = T + B[i]
A[i + 1] = T
The pass relies on loop dependence analysis via LoopAccessAnalisys to
find opportunities of loop-carried dependences with a distance of one
between a store and a load. Since it's using LoopAccessAnalysis, it was
easy to also add support for versioning away may-aliasing intervening
stores that would otherwise prevent this transformation.
This optimization is also performed by Load-PRE in GVN without the
option of multi-versioning. As was discussed with Daniel Berlin in
http://reviews.llvm.org/D9548, this is inferior to a more loop-aware
solution applied here. Hopefully, we will be able to remove some
complexity from GVN/MemorySSA as a consequence.
In the long run, we may want to extend this pass (or create a new one if
there is little overlap) to also eliminate loop-indepedent redundant
loads and store that *require* versioning due to may-aliasing
intervening stores/loads. I have some motivating cases for store
elimination. My plan right now is to wait for MemorySSA to come online
first rather than using memdep for this.
The main motiviation for this pass is the 456.hmmer loop in SPECint2006
where after distributing the original loop and vectorizing the top part,
we are left with the critical path exposed in the bottom loop. Being
able to promote the memory dependence into a register depedence (even
though the HW does perform store-to-load fowarding as well) results in a
major gain (~20%). This gain also transfers over to x86: it's
around 8-10%.
Right now the pass is off by default and can be enabled
with -enable-loop-load-elim. On the LNT testsuite, there are two
performance changes (negative number -> improvement):
1. -28% in Polybench/linear-algebra/solvers/dynprog: the length of the
critical paths is reduced
2. +2% in Polybench/stencils/adi: Unfortunately, I couldn't reproduce this
outside of LNT
The pass is scheduled after the loop vectorizer (which is after loop
distribution). The rational is to try to reuse LAA state, rather than
recomputing it. The order between LV and LLE is not critical because
normally LV does not touch scalar st->ld forwarding cases where
vectorizing would inhibit the CPU's st->ld forwarding to kick in.
LoopLoadElimination requires LAA to provide the full set of dependences
(including forward dependences). LAA is known to omit loop-independent
dependences in certain situations. The big comment before
removeDependencesFromMultipleStores explains why this should not occur
for the cases that we're interested in.
Reviewers: dberlin, hfinkel
Subscribers: junbuml, dberlin, mssimpso, rengolin, sanjoy, llvm-commits
Differential Revision: http://reviews.llvm.org/D13259
llvm-svn: 252017
Now that all the known faults with GlobalsAA have been fixed, flip the big switch on -enable-non-lto-gmr again.
Feel free to pester me with any more bugs found, and don't hesitate to flip the switch back off.
llvm-svn: 250157
Loop unswitching produces conditional branches with constant condition,
and it's beneficial for later passes to clean this up with simplify-cfg.
We do this after the second invocation of loop-unswitch, but not after
the first one. Not doing so might cause problem for passes like
LoopUnroll, whose estimate of loop body size would be less accurate.
Reviewers: hfinkel
Differential Revision: http://reviews.llvm.org/D13064
llvm-svn: 248460
David Majnemer