Summary:
This is the first set of changes implementing the RFC from
http://thread.gmane.org/gmane.comp.compilers.llvm.devel/98334
This is a cross-sectional patch; rather than implementing the hotness
attribute for all optimization remarks and all passes in a patch set, it
implements it for the 'missed-optimization' remark for Loop
Distribution. My goal is to shake out the design issues before scaling
it up to other types and passes.
Hotness is computed as an integer as the multiplication of the block
frequency with the function entry count. It's only printed in opt
currently since clang prints the diagnostic fields directly. E.g.:
remark: /tmp/t.c:3:3: loop not distributed: use -Rpass-analysis=loop-distribute for more info (hotness: 300)
A new API added is similar to emitOptimizationRemarkMissed. The
difference is that it additionally takes a code region that the
diagnostic corresponds to. From this, hotness is computed using BFI.
The new API is exposed via an analysis pass so that it can be made
dependent on LazyBFI. (Thanks to Hal for the analysis pass idea.)
This feature can all be enabled by setDiagnosticHotnessRequested in the
LLVM context. If this is off, LazyBFI is not calculated (D22141) so
there should be no overhead.
A new command-line option is added to turn this on in opt.
My plan is to switch all user of emitOptimizationRemark* to use this
module instead.
Reviewers: hfinkel
Subscribers: rcox2, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D21771
llvm-svn: 275583
Summary: Convert LoopInstSimplify to new PM. Unfortunately there is no exisiting unittest for this pass.
Reviewers: davidxl, silvas
Subscribers: silvas, llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D22280
llvm-svn: 275576
Summary:
This change use the overlap interval map built from partial overwrite tracking to perform shortening MemIntrinsics.
Add test cases which was missing opportunities before.
Reviewers: hfinkel, eeckstein, mcrosier
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D21909
llvm-svn: 275571
This pass hoists duplicated computations in the program. The primary goal of
gvn-hoist is to reduce the size of functions before inline heuristics to reduce
the total cost of function inlining.
Pass written by Sebastian Pop, Aditya Kumar, Xiaoyu Hu, and Brian Rzycki.
Important algorithmic contributions by Daniel Berlin under the form of reviews.
Differential Revision: http://reviews.llvm.org/D19338
llvm-svn: 275561
Fix for PR 28418.
opt never finishes compiling a test when -gvn option is passed.
The problem is caused by the fact that GVN fails to fold a constant expression.
Differential Revision: https://reviews.llvm.org/D22185
llvm-svn: 275483
Summary: Port Dead Loop Deletion Pass to the new pass manager.
Reviewers: silvas, davide
Subscribers: llvm-commits, sanjoy, mcrosier
Differential Revision: https://reviews.llvm.org/D21483
llvm-svn: 275453
enables the code size optimisation to fold a rem and div into a single
aeabi_uidivmod call. This was not happening before because sdiv was converted
but srem not, and instructions with different signedness are not combined.
Differential Revision: http://reviews.llvm.org/D22214
llvm-svn: 275403
This pass hoists duplicated computations in the program. The primary goal of
gvn-hoist is to reduce the size of functions before inline heuristics to reduce
the total cost of function inlining.
Pass written by Sebastian Pop, Aditya Kumar, Xiaoyu Hu, and Brian Rzycki.
Important algorithmic contributions by Daniel Berlin under the form of reviews.
Differential Revision: http://reviews.llvm.org/D19338
llvm-svn: 275401
constant hoisting. It not only takes into account the number of uses and the
cost of expressions in which constants appear, but now also the resulting
integer range of the offsets. Thus, the algorithm maximizes the number of uses
within an integer range that will enable more efficient code generation. On
ARM, for example, this will enable code size optimisations because less
negative offsets will be created. Negative offsets/immediates are not supported
by Thumb1 thus preventing more compact instruction encoding.
Differential Revision: http://reviews.llvm.org/D21183
llvm-svn: 275382
This now should also work with the interprocedural variant of the pass.
Slightly easier now that the yak is shaved.
Differential Revision: http://reviews.llvm.org/D22329
llvm-svn: 275363
Summary:
In Scalarizer::gather we see if we already have a scattered form of Op,
and in that case use the new form.
In the particular case of PR28108, the found ValueVector SV has size 2,
where the first Value is nullptr, and the second is indeed a proper Value.
The nullptr then caused an assert to blow when we tried to do
cast<Instruction>(SV[I]).
With this patch we check SV[I] before doing the cast, and if it's nullptr
we just skip over it.
I don't know the Scalarizer well enough to know if this is the best fix
or if something should be done else where to prevent the nullptr from
being in the ValueVector at all, but at least this avoids the crash
and looking at the test case output it looks reasonable.
Reviewers: hfinkel, frasercrmck, wala, mehdi_amini
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D21518
llvm-svn: 275359
The code was pretty much copy-pasted between SCCP and IPSCCP. The situation
became clearly worse after I introduced the support for folding structs in
SCCP. This commit is NFC as we currently (still) skip the replacement
step in IPSCCP, but I'll change this soon.
llvm-svn: 275339
In the solver, isUndefined() does really mean "we don't know the
value yet" rather than "this is an UndefinedValue". Discussed with
Eli Friedman.
Differential Revision: http://reviews.llvm.org/D22192
llvm-svn: 275004
This code was already commented out and it made some weird assumptions,
e.g. using isUndefined() as "this value is UndefValue" instead of
"we haven't computed this value is yet". Thanks to Eli Friedman for
pointing out where I was wrong (and where this code was wrong).
llvm-svn: 274995
ConstantInt::getSExtValue may fail on >64-bit integers. Add checks to call
getSExtValue only on narrow integers.
As a minor aside, simplify slsr-gep.ll to remove unnecessary load instructions.
llvm-svn: 274982
This should be slightly more efficient and could avoid spurious overdefined
markings, as Eli pointed out.
Differential Revision: http://reviews.llvm.org/D22122
llvm-svn: 274905
We can remove dead stores in the presence of fence instructions. Fence
does not change an otherwise thread local store to visible.
reviewers: reames, dexonsmith, jfb
Differential Revision: http://reviews.llvm.org/D22001
llvm-svn: 274795
The dse_with_dbg_value.ll test committed with r273141 is removed because this
we no longer performs any type of back tracking, which is what was causing the
codegen differences with and without debug information.
Differential Revision: http://reviews.llvm.org/D21613
llvm-svn: 274660
We were still crashing in the "no change" case because LVI was not
getting invalidated.
See the thread "Should analyses be able to hold AssertingVH to IR?
(related to PR28400)" for more discussion.
llvm-svn: 274656
PR28400 seems to be not an isolated issue, but a general problem related
to caching analyses. We will need to discuss on llvm-dev.
A test case is in the PR.
llvm-svn: 274457
This actually uncovered a surprisingly large chain of ultimately unused
TLI args.
From what I can gather, this argument is a remnant of when
isKnownNonNull would look at the TLI directly.
The current approach seems to be that InferFunctionAttrs runs early in
the pipeline and uses TLI to annotate the TLI-dependent non-null
information as return attributes.
This also removes the dependence of functionattrs on TLI altogether.
llvm-svn: 274455
This pass hoists duplicated computations in the program. The primary goal of
gvn-hoist is to reduce the size of functions before inline heuristics to reduce
the total cost of function inlining.
Pass written by Sebastian Pop, Aditya Kumar, Xiaoyu Hu, and Brian Rzycki.
Important algorithmic contributions by Daniel Berlin under the form of reviews.
Differential Revision: http://reviews.llvm.org/D19338
llvm-svn: 274305
Summary:
Found cases where DSE incorrectly add partially-overwritten intervals.
Please see the test case for details.
Reviewers: mcrosier, eeckstein, hfinkel
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D21859
llvm-svn: 274237
For the new hotness attribute, the API will take the pass rather than
the pass name so we can no longer play the trick of AlwaysPrint being a
special pass name. This adds a getter to help the transition.
There is also a corresponding clang patch.
llvm-svn: 274100
If the load is conditional we can't hoist its 0-iteration instance to
the preheader because that would make it unconditional. Thus we would
access a memory location that the original loop did not access.
llvm-svn: 273991
Summary:
This is a straightforward extension of what LoopUnswitch does to
branches to guards. That is, we unswitch
```
for (;;) {
...
guard(loop_invariant_cond);
...
}
```
into
```
if (loop_invariant_cond) {
for (;;) {
...
// There is no need to emit guard(true)
...
}
} else {
for (;;) {
...
guard(false);
// SimplifyCFG will clean this up by adding an
// unreachable after the guard(false)
...
}
}
```
Reviewers: majnemer
Subscribers: mcrosier, llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D21725
llvm-svn: 273801
SimplifyCFG had logic to insert calls to llvm.trap for two very
particular IR patterns: stores and invokes of undef/null.
While InstCombine canonicalizes certain undefined behavior IR patterns
to stores of undef, phase ordering means that this cannot be relied upon
in general.
There are much better tools than llvm.trap: UBSan and ASan.
N.B. I could be argued into reverting this change if a clear argument as
to why it is important that we synthesize llvm.trap for stores, I'd be
hard pressed to see why it'd be useful for invokes...
llvm-svn: 273778
r273711 was reverted by r273743. The inliner needs to know about any
call sites in the inlined function. These were obscured if we replaced
a call to undef with an undef but kept the call around.
This fixes PR28298.
llvm-svn: 273753
Summary:
We can avoid repeating the check `isGuaranteedToExecute` when it's already called once while checking if the alignment can be widened for the load/store being hoisted.
The function is invariant for the same instruction `UI` in `isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo);`
Reviewers: hfinkel, eli.friedman
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D21672
llvm-svn: 273671
DeadStoreElimination can currently remove a small store rendered unnecessary by
a later larger one, but could not remove a larger store rendered unnecessary by
a series of later smaller ones. This adds that capability.
It works by keeping a map, which is used as an effective interval map, for each
store later overwritten only partially, and filling in that interval map as
more such stores are discovered. No additional walking or aliasing queries are
used. In the map forms an interval covering the the entire earlier store, then
it is dead and can be removed. The map is used as an interval map by storing a
mapping between the ending offset and the beginning offset of each interval.
I discovered this problem when investigating a performance issue with code like
this on PowerPC:
#include <complex>
using namespace std;
complex<float> bar(complex<float> C);
complex<float> foo(complex<float> C) {
return bar(C)*C;
}
which produces this:
define void @_Z4testSt7complexIfE(%"struct.std::complex"* noalias nocapture sret %agg.result, i64 %c.coerce) {
entry:
%ref.tmp = alloca i64, align 8
%tmpcast = bitcast i64* %ref.tmp to %"struct.std::complex"*
%c.sroa.0.0.extract.shift = lshr i64 %c.coerce, 32
%c.sroa.0.0.extract.trunc = trunc i64 %c.sroa.0.0.extract.shift to i32
%0 = bitcast i32 %c.sroa.0.0.extract.trunc to float
%c.sroa.2.0.extract.trunc = trunc i64 %c.coerce to i32
%1 = bitcast i32 %c.sroa.2.0.extract.trunc to float
call void @_Z3barSt7complexIfE(%"struct.std::complex"* nonnull sret %tmpcast, i64 %c.coerce)
%2 = bitcast %"struct.std::complex"* %agg.result to i64*
%3 = load i64, i64* %ref.tmp, align 8
store i64 %3, i64* %2, align 4 ; <--- ***** THIS SHOULD NOT BE HERE ****
%_M_value.realp.i.i = getelementptr inbounds %"struct.std::complex", %"struct.std::complex"* %agg.result, i64 0, i32 0, i32 0
%4 = lshr i64 %3, 32
%5 = trunc i64 %4 to i32
%6 = bitcast i32 %5 to float
%_M_value.imagp.i.i = getelementptr inbounds %"struct.std::complex", %"struct.std::complex"* %agg.result, i64 0, i32 0, i32 1
%7 = trunc i64 %3 to i32
%8 = bitcast i32 %7 to float
%mul_ad.i.i = fmul fast float %6, %1
%mul_bc.i.i = fmul fast float %8, %0
%mul_i.i.i = fadd fast float %mul_ad.i.i, %mul_bc.i.i
%mul_ac.i.i = fmul fast float %6, %0
%mul_bd.i.i = fmul fast float %8, %1
%mul_r.i.i = fsub fast float %mul_ac.i.i, %mul_bd.i.i
store float %mul_r.i.i, float* %_M_value.realp.i.i, align 4
store float %mul_i.i.i, float* %_M_value.imagp.i.i, align 4
ret void
}
the problem here is not just that the i64 store is unnecessary, but also that
it blocks further backend optimizations of the other uses of that i64 value in
the backend.
In the future, we might want to add a special case for handling smaller
accesses (e.g. using a bit vector) if the map mechanism turns out to be
noticeably inefficient. A sorted vector is also a possible replacement for the
map for small numbers of tracked intervals.
Differential Revision: http://reviews.llvm.org/D18586
llvm-svn: 273559
This reverts commit r273160, reapplying r273132.
RecursivelyDeleteTriviallyDeadInstructions cannot be called on a
parentless Instruction.
llvm-svn: 273162
This reverts commit r273132.
Breaks multiple test under /llvm/test:Transforms (e.g.
llvm/test:Transforms/LoopIdiom/basic.ll.test) under asan.
llvm-svn: 273160
After a store has been eliminated, when making sure that the
instruction iterator points to a valid instruction, dbg intrinsics are
now ignored as a new instruction.
Patch by Henric Karlsson.
Reviewed by Daniel Berlin.
Differential Revision: http://reviews.llvm.org/D21076
llvm-svn: 273141
Removing dead instructions requires remembering which operands have
already been removed. RecursivelyDeleteTriviallyDeadInstructions has
this logic, don't partially reimplement it in LoopIdiomRecognize.
This fixes PR28196.
llvm-svn: 273132
Change the underlying offset and comparisons to use int64_t instead of
uint64_t.
Patch by River Riddle!
Differential Revision: http://reviews.llvm.org/D21499
llvm-svn: 273105
This is a functional change for LLE and LDist. The other clients (LV,
LVerLICM) already had this explicitly enabled.
The temporary boolean parameter to LAA is removed that allowed turning
off speculation of symbolic strides. This makes LAA's caching interface
LAA::getInfo only take the loop as the parameter. This makes the
interface more friendly to the new Pass Manager.
The flag -enable-mem-access-versioning is moved from LV to a LAA which
now allows turning off speculation globally.
llvm-svn: 273064
This is indeed a much cleaner approach (thanks to Daniel Berlin
for pointing out), and also David/Sean for review.
Differential Revision: http://reviews.llvm.org/D21454
llvm-svn: 273032
pass manager passes' `run` methods.
This removes a bunch of SFINAE goop from the pass manager and just
requires pass authors to accept `AnalysisManager<IRUnitT> &` as a dead
argument. This is a small price to pay for the simplicity of the system
as a whole, despite the noise that changing it causes at this stage.
This will also helpfull allow us to make the signature of the run
methods much more flexible for different kinds af passes to support
things like intelligently updating the pass's progression over IR units.
While this touches many, many, files, the changes are really boring.
Mostly made with the help of my trusty perl one liners.
Thanks to Sean and Hal for bouncing ideas for this with me in IRC.
llvm-svn: 272978
This is still NFCI, so the list of clients that allow symbolic stride
speculation does not change (yes: LV and LoopVersioningLICM, no: LLE,
LDist). However since the symbolic strides are now managed by LAA
rather than passed by client a new bool parameter is used to enable
symbolic stride speculation.
The existing test Transforms/LoopVectorize/version-mem-access.ll checks
that stride speculation is performed for LV.
The previously added test Transforms/LoopLoadElim/symbolic-stride.ll
ensures that no speculation is performed for LLE.
The next patch will change the functionality and turn on symbolic stride
speculation in all of LAA's clients and remove the bool parameter.
llvm-svn: 272970
Redundant invariant loads can be CSE'ed with very little extra effort
over what early-cse already tracks, so it looks reasonable to make
early-cse handle this case.
llvm-svn: 272954
Daniel Berlin expressed some real concerns about the port and proposed
and alternative approach. I'll revert this for now while working on a
new patch, which I hope to put up for review shortly. Sorry for the churn.
llvm-svn: 272925
We should update results of the BranchProbabilityInfo after removing block in JumpThreading. Otherwise
we will get dangling pointer inside BranchProbabilityInfo cache.
Differential Revision: http://reviews.llvm.org/D20957
llvm-svn: 272891
Added checks to make sure the Scalarizer::transferMetadata() don't
remove valid debug locations from instructions. This is important as
the verifier pass require that e.g. inlinable callsites have a valid
debug location.
https://llvm.org/bugs/show_bug.cgi?id=27938
Patch by Karl-Johan Karlsson
Reviewers: dblaikie
Differential Revision: http://reviews.llvm.org/D20807
llvm-svn: 272884
We do not support splitting cleanuppad or catchswitches. This is
problematic for passes which assume that a loop is in loop simplify
form (the loop would have a dedicated exit block instead of sharing it).
While it isn't great that we don't support this for cleanups, we still
cannot make loop-simplify form an assertable precondition because
indirectbr will also disable these sorts of CFG cleanups.
This fixes PR28132.
llvm-svn: 272739
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
If a local_unnamed_addr attribute is attached to a global, the address
is known to be insignificant within the module. It is distinct from the
existing unnamed_addr attribute in that it only describes a local property
of the module rather than a global property of the symbol.
This attribute is intended to be used by the code generator and LTO to allow
the linker to decide whether the global needs to be in the symbol table. It is
possible to exclude a global from the symbol table if three things are true:
- This attribute is present on every instance of the global (which means that
the normal rule that the global must have a unique address can be broken without
being observable by the program by performing comparisons against the global's
address)
- The global has linkonce_odr linkage (which means that each linkage unit must have
its own copy of the global if it requires one, and the copy in each linkage unit
must be the same)
- It is a constant or a function (which means that the program cannot observe that
the unique-address rule has been broken by writing to the global)
Although this attribute could in principle be computed from the module
contents, LTO clients (i.e. linkers) will normally need to be able to compute
this property as part of symbol resolution, and it would be inefficient to
materialize every module just to compute it.
See:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160509/356401.htmlhttp://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160516/356738.html
for earlier discussion.
Part of the fix for PR27553.
Differential Revision: http://reviews.llvm.org/D20348
llvm-svn: 272709
We move the loop rotate functions in a separate class to avoid passing multiple
parameters to each function. This cleanup will help with further development of
loop rotation. NFC.
Patch written by Aditya Kumar and Sebastian Pop.
Differential Revision: http://reviews.llvm.org/D21311
llvm-svn: 272672
The need for all these Lookup* functions is just because of calls to
getAnalysis inside methods (i.e. not at the top level) of the
runOnFunction method. They should be straightforward to clean up when
the old PM is gone.
llvm-svn: 272615
This reverts commit r272603 and adds a fix.
Big thanks to Davide for pointing me at r216244 which gives some insight
into how to fix this VS2013 issue. VS2013 can't synthesize a move
constructor. So the fix here is to add one explicitly to the
JumpThreadingPass class.
llvm-svn: 272607
This follows the approach in r263208 (for GVN) pretty closely:
- move the bulk of the body of the function to the new PM class.
- expose a runImpl method on the new-PM class that takes the IRUnitT and
pointers/references to any analyses and use that to implement the
old-PM class.
- use a private namespace in the header for stuff that used to be file
scope
llvm-svn: 272597
This is a bit gnarly since LVI is maintaining its own cache.
I think this port could be somewhat cleaner, but I'd rather not spend
too much time on it while we still have the old pass hanging around and
limiting how much we can clean things up.
Once the old pass is gone it will be easier (less time spent) to clean
it up anyway.
This is the last dependency needed for porting JumpThreading which I'll
do in a follow-up commit (there's no printer pass for LVI or anything to
test it, so porting a pass that depends on it seems best).
I've been mostly following:
r269370 / D18834 which ported Dependence Analysis
r268601 / D19839 which ported BPI
llvm-svn: 272593
This used to be free, copying and moving DebugLocs became expensive
after the metadata rewrite. Passing by reference eliminates a ton of
track/untrack operations. No functionality change intended.
llvm-svn: 272512
It isn't legal to hoist a load past a call which might not return;
even if it doesn't throw, it could, for example, call exit().
Fixes http://llvm.org/PR27953.
llvm-svn: 272495
Summary:
Make isGuaranteedToExecute use the
isGuaranteedToTransferExecutionToSuccessor helper, and make that helper
a bit more accurate.
There's a potential performance impact here from assuming that arbitrary
calls might not return. This probably has little impact on loads and
stores to a pointer because most things alias analysis can reason about
are dereferenceable anyway. The other impacts, like less aggressive
hoisting of sdiv by a variable and less aggressive hoisting around
volatile memory operations, are unlikely to matter for real code.
This also impacts SCEV, which uses the same helper. It's a minor
improvement there because we can tell that, for example, memcpy always
returns normally. Strictly speaking, it's also introducing
a bug, but it's not any worse than everywhere else we assume readonly
functions terminate.
Fixes http://llvm.org/PR27857.
Reviewers: hfinkel, reames, chandlerc, sanjoy
Subscribers: broune, llvm-commits
Differential Revision: http://reviews.llvm.org/D21167
llvm-svn: 272489
Summary:
I can't find a case where we can rotate a loop more than once, and it looks
like we never do this. To rotate a loop following conditions should be met:
1) its header should be exiting
2) its latch shouldn't be exiting
But after the first rotation the header becomes the new latch, so this
condition can never be true any longer.
Tested on with an assert on LNT testsuite and make check.
Reviewers: hfinkel, sanjoy
Subscribers: sebpop, sanjoy, llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D20181
llvm-svn: 272439
Loop unswitching may cause MSan false positive when the unswitch
condition is not guaranteed to execute.
This is very similar to ASan and TSan special case in
llvm::isSafeToSpeculativelyExecute (they don't like speculative loads
and stores), but for branch instructions.
This is a workaround for PR28054.
llvm-svn: 272421
with user specified count has been applied.
Summary:
Previously SetLoopAlreadyUnrolled() set the disable pragma only if
there was some loop metadata.
Now it set the pragma in all cases. This helps to prevent multiple
unroll when -unroll-count=N is given.
Reviewers: mzolotukhin
Differential Revision: http://reviews.llvm.org/D20765
From: Evgeny Stupachenko <evstupac@gmail.com>
llvm-svn: 272195
As suggested by clang-tidy's performance-unnecessary-copy-initialization.
This can easily hit lifetime issues, so I audited every change and ran the
tests under asan, which came back clean.
llvm-svn: 272126
Summary:
Fix LSRInstance::HoistInsertPosition() to check the original insert
position block first for a canonical insertion point that is dominated
by all inputs. This leads to SCEV being able to reuse more instructions
since it currently tracks the instructions it creates for reuse by
keeping a table of <Value, insert point> pairs.
Originally reviewed in http://reviews.llvm.org/D18001
Reviewers: atrick
Subscribers: llvm-commits, mzolotukhin, mcrosier
Differential Revision: http://reviews.llvm.org/D18480
llvm-svn: 271929
Summary:
This hasn't been caught before because it requires noalias or similarly
strong alias analysis to actually reproduce.
Fixes http://llvm.org/PR27952 .
Reviewers: hfinkel, sanjoy
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D20944
llvm-svn: 271858
Summary:
There are some rough corners, since the new pass manager doesn't have
(as far as I can tell) LoopSimplify and LCSSA, so I've updated the
tests to run them separately in the old pass manager in the lit tests.
We also don't have an equivalent for AU.setPreservesCFG() in the new
pass manager, so I've left a FIXME.
Reviewers: bogner, chandlerc, davide
Subscribers: sanjoy, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D20783
llvm-svn: 271846
In r270478, where I enabled the new heuristic I posted testing results,
which I got when explicitly passed the thresholds values via CL options.
However, setting the CL options init-values is not enough to change the
default values of thresholds, so I'm changing them in another place now.
llvm-svn: 271615
In preparation for porting to the new PM.
Patch by Jake VanAdrighem! (review mainly by me/Justin)
Differential Revision: http://reviews.llvm.org/D20610
llvm-svn: 271607
Since we already assert that the outgoing IR is in LCSSA, it is easy to
get misled into thinking that -indvars broke LCSSA if the incoming IR is
non-LCSSA. Checking this pre-condition will make such cases break in
more obvious ways.
Inspired by (but does _not_ fix) PR26682.
llvm-svn: 271196
Summary:
Unroll factor (Count) calculations moved to a new function.
Early exits on pragma and "-unroll-count" defined factor added.
New type of unrolling "Force" introduced (previously used implicitly).
New unroll preference "AllowRemainder" introduced and set "true" by default.
(should be set to false for architectures that suffers from it).
Reviewers: hfinkel, mzolotukhin, zzheng
Differential Revision: http://reviews.llvm.org/D19553
From: Evgeny Stupachenko <evstupac@gmail.com>
llvm-svn: 271071
Currently we consider that each constant has itself as a base value. I.e "base(const) = const".
This introduces couple of problems when we are trying to avoid reporting constants in statepoint live sets:
1. When querying "base( phi(const1, const2) )" we will get "phi(const1, const2)" as a base pointer. Since
it's not a constant we will record it in a stack map. However on practice we don't want this to happen
(constant are never relocated).
2. base( phi(const, gc ptr) ) = phi( const, base(gc ptr) ). This particular case imposes challenge on our
runtime - we don't expect to see constant base pointers other than null. This problems can be avoided
by treating all constant as if they were derived from null pointer base. I.e in a first case we will
not include constant pointer in a stack map at all. In a second case we will get "phi(null, base(gc ptr))"
as a base pointer which is a lot more convenient.
Differential Revision: http://reviews.llvm.org/D20584
llvm-svn: 270993
Condition might be simplified to a Constant, but it doesn't have to be
ConstantInt, so we should dyn_cast, instead of cast.
This fixes PR27886.
llvm-svn: 270924
An exception could prevent a store from occurring but MemCpyOpt's
callslot optimization would fire anyway, causing the store to occur.
This fixes PR27849.
llvm-svn: 270892
It is unsafe to hoist a load before a function call which may throw, the
throw might prevent a pointer dereference.
Likewise, it is unsafe to sink a store after a call which may throw.
The caller might be able to observe the difference.
This fixes PR27858.
llvm-svn: 270828
After this change, we do the expected thing for cases like
```
Check0Passed = /* range check IRCE can optimize */
Check1Passed = /* range check IRCE can optimize */
if (!(Check0Passed && Check1Passed))
throw_Exception();
```
llvm-svn: 270804
This changes IRCE to optimize uses, and not branches. This change is
NFCI since the uses we do inspect are in practice only ever going to be
the condition use in conditional branches; but this flexibility will
later allow us to analyze more complex expressions than just a direct
branch on a range check.
llvm-svn: 270500
Summary:
This patch turns on LoopUnrollAnalyzer by default. To mitigate compile
time regressions, I chose very conservative thresholds for now. Later we
can make them more aggressive, but it might require being smarter in
which loops we're optimizing. E.g. currently the biggest issue is that
with more agressive thresholds we unroll many cold loops, which
increases compile time for no performance benefit (performance of those
loops is improved, but it doesn't matter since they are cold).
Test results for compile time(using 4 samples to reduce noise):
```
MultiSource/Benchmarks/VersaBench/ecbdes/ecbdes 5.19%
SingleSource/Benchmarks/Polybench/medley/reg_detect/reg_detect 4.19%
MultiSource/Benchmarks/FreeBench/fourinarow/fourinarow 3.39%
MultiSource/Applications/JM/lencod/lencod 1.47%
MultiSource/Benchmarks/Fhourstones-3_1/fhourstones3_1 -6.06%
```
I didn't see any performance changes in the testsuite, but it improves
some internal tests.
Reviewers: hfinkel, chandlerc
Subscribers: llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D20482
llvm-svn: 270478
The InductiveRangeCheck struct is only five words long; so passing these
around value is fine. The allocator makes the code look more complex
than it is.
llvm-svn: 270309
I had used `std::remove_if` under the assumption that it moves the
predicate matching elements to the end, but actaully the elements
remaining towards the end (after the iterator returned by
`std::remove_if`) are indeterminate. Fix the bug (and make the code
more straightforward) by using a temporary SmallVector, and add a test
case demonstrating the issue.
llvm-svn: 270306
Sequences of range checks expressed using guards, like
guard((I - 2) u< L)
guard((I - 1) u< L)
guard((I + 0) u< L)
guard((I + 1) u< L)
guard((I + 2) u< L)
can sometimes be combined into a smaller sequence:
guard((I - 2) u< L AND (I + 2) u< L)
if we can prove that (I - 2) u< L AND (I + 2) u< L implies all of checks
expressed in the previous sequence.
This change teaches GuardWidening to do this kind of merging when
feasible.
llvm-svn: 270151
Summary:
Implement guard widening in LLVM. Description from GuardWidening.cpp:
The semantics of the `@llvm.experimental.guard` intrinsic lets LLVM
transform it so that it fails more often that it did before the
transform. This optimization is called "widening" and can be used hoist
and common runtime checks in situations like these:
```
%cmp0 = 7 u< Length
call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
call @unknown_side_effects()
%cmp1 = 9 u< Length
call @llvm.experimental.guard(i1 %cmp1) [ "deopt"(...) ]
...
```
to
```
%cmp0 = 9 u< Length
call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
call @unknown_side_effects()
...
```
If `%cmp0` is false, `@llvm.experimental.guard` will "deoptimize" back
to a generic implementation of the same function, which will have the
correct semantics from that point onward. It is always _legal_ to
deoptimize (so replacing `%cmp0` with false is "correct"), though it may
not always be profitable to do so.
NB! This pass is a work in progress. It hasn't been tuned to be
"production ready" yet. It is known to have quadriatic running time and
will not scale to large numbers of guards
Reviewers: reames, atrick, bogner, apilipenko, nlewycky
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D20143
llvm-svn: 269997
This is assertion is no longer necessary since we never record
constants in the live set anyway. (They are never recorded in
the initial live set, and constant bases are removed near line 2119)
Differential Revision: http://reviews.llvm.org/D20293
llvm-svn: 269764
TargetLibraryInfoWrapperPass is a dependency of
SCCP but it's not listed as such. Chandler pointed
out this is an easy mistake to make which only
surfaces in weird crashes with some flag combinations.
This code will go away anyway at some point in the
future, but as long as it's (still) exercised, try
to make it correct.
llvm-svn: 269589
Summary: This change fix the bug in isProfitableToUseMemset() where MaxIntSize shoule be in byte, not bit.
Reviewers: arsenm, joker.eph, mcrosier
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D20176
llvm-svn: 269433
Summary:
...loop after the last iteration.
This is really hard to do correctly. The core problem is that we need to
model liveness through the induction PHIs from iteration to iteration in
order to get the correct results, and we need to correctly de-duplicate
the common subgraphs of instructions feeding some subset of the
induction PHIs. All of this can be driven either from a side effect at
some iteration or from the loop values used after the loop finishes.
This patch implements this by storing the forward-propagating analysis
of each instruction in a cache to recall whether it was free and whether
it has become live and thus counted toward the total unroll cost. Then,
at each sink for a value in the loop, we recursively walk back through
every value that feeds the sink, including looping back through the
iterations as needed, until we have marked the entire input graph as
live. Because we cache this, we never visit instructions more than twice
-- once when we analyze them and put them into the cache, and once when
we count their cost towards the unrolled loop. Also, because the cache
is only two bits and because we are dealing with relatively small
iteration counts, we can store all of this very densely in memory to
avoid this from becoming an excessively slow analysis.
The code here is still pretty gross. I would appreciate suggestions
about better ways to factor or split this up, I've stared too long at
the algorithmic side to really have a good sense of what the design
should probably look at.
Also, it might seem like we should do all of this bottom-up, but I think
that is a red herring. Specifically, the simplification power is *much*
greater working top-down. We can forward propagate very effectively,
even across strange and interesting recurrances around the backedge.
Because we use data to propagate, this doesn't cause a state space
explosion. Doing this level of constant folding, etc, would be very
expensive to do bottom-up because it wouldn't be until the last moment
that you could collapse everything. The current solution is essentially
a top-down simplification with a bottom-up cost accounting which seems
to get the best of both worlds. It makes the simplification incremental
and powerful while leaving everything dead until we *know* it is needed.
Finally, a core property of this approach is its *monotonicity*. At all
times, the current UnrolledCost is a conservatively low estimate. This
ensures that we will never early-exit from the analysis due to exceeding
a threshold when if we had continued, the cost would have gone back
below the threshold. These kinds of bugs can cause incredibly hard to
track down random changes to behavior.
We could use a techinque similar (but much simpler) within the inliner
as well to avoid considering speculated code in the inline cost.
Reviewers: chandlerc
Subscribers: sanjoy, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D11758
llvm-svn: 269388
Ported DA to the new PM by splitting the former DependenceAnalysis Pass
into a DependenceInfo result type and DependenceAnalysisWrapperPass type
and adding a new PM-style DependenceAnalysis analysis pass returning the
DependenceInfo.
Patch by Philip Pfaffe, most of the review by Justin.
Differential Revision: http://reviews.llvm.org/D18834
llvm-svn: 269370
Shifts beyond the bitwidth are undef but SCCP resolved them to zero.
Instead, DTRT and resolve them to undef.
This reimplements the transform which caused PR27712.
llvm-svn: 269269
Sort of the BB-local equivalent to idiom-recognizer: if we have a basic-block
that really implements a memcpy operation, backends can benefit from seeing
this.
llvm-svn: 269125
Before r268509, Clang would disable the loop unroll pass when optimizing
for size. That commit enabled it to be able to support unroll pragmas
in -Os builds. However, this regressed binary size in one of Chromium's
DLLs with ~100 KB.
This restores the original behaviour of no unrolling at -Os, but doing it
in LLVM instead of Clang makes more sense, and also allows the pragmas to
keep working.
Differential revision: http://reviews.llvm.org/D20115
llvm-svn: 269124
This patch extend loopreroll to allow the instruction chain
of loop control only IV has sext.
Differential Revision: http://reviews.llvm.org/D19820
llvm-svn: 269121
Loop rotation clones instruction from the old header into the preheader. If
there were uses of values produced by these instructions that were outside
the loop, we have to insert PHI nodes to merge the two values. If the values
are used by DbgIntrinsics they will be used as a MetadataAsValue of a
ValueAsMetadata of the original values, and iterating all of the uses of the
original value will not update the DbgIntrinsics. The new code checks if the
values are used by DbgIntrinsics and if so, updates them using essentially
the same logic as the original code.
The attached testcase demonstrates the issue. Without the fix, the
DbgIntrinic outside the loop uses values computed inside the loop, even
though these values do not dominate the DbgIntrinsic.
Author: Thomas Jablin (tjablin)
Reviewers: dblaikie aprantl kbarton hfinkel cycheng
http://reviews.llvm.org/D19564
llvm-svn: 269034
Again, fairly simple. Only change is ensuring that we actually copy the property of the load correctly. The aliasing legality constraints were already handled by the FRE patches. There's nothing special about unorder atomics from the perspective of the PRE algorithm itself.
llvm-svn: 268804
You'll note there are essentially no code changes here. Cross block FRE heavily reuses code from the block local FRE. All of the tricky parts were done as part of the previous patch and the refactoring that removed the original code duplication.
llvm-svn: 268775
This patch is the first in a small series teaching GVN to optimize unordered loads aggressively. This change just handles block local FRE because that's the simplest thing which lets me test MDA, and the AvailableValue pieces. Somewhat suprisingly, MDA appears fine and only a couple of small changes are needed in GVN.
Once this is in, I'll tackle non-local FRE and PRE. The former looks like a natural extension of this, the later will require a couple of minor changes.
Differential Revision: http://reviews.llvm.org/D19440
llvm-svn: 268770
Summary: We need to clean up CFG before assigning discriminator to minimize the impact of optimization on debug info.
Reviewers: davidxl, dblaikie, dnovillo
Subscribers: dnovillo, danielcdh, llvm-commits
Differential Revision: http://reviews.llvm.org/D19926
llvm-svn: 268675
Goal of this change is to guarantee stable ordering of the statepoint arguments and other
newly inserted values such as gc.relocates. Previously we had explicit sorting in a couple
of places. However for unnamed values ordering was partial and overall we didn't have any
strong invariant regarding it. This change switches all data structures to use SetVector's
and MapVector's which provide possibility for deterministic iteration over them.
Explicit sorting is now redundant and was removed.
Differential Revision: http://reviews.llvm.org/D19669
llvm-svn: 268502
pointing to the same addr space. This can prevent SROA from creating a bitcast
between pointers with different addr spaces.
Differential Revision: http://reviews.llvm.org/D19697
llvm-svn: 268424
SCEV caches whether SCEV expressions are loop invariant, variant or
computable. LICM breaks this cache, almost by definition; so clear the
SCEV disposition cache if LICM changed anything.
llvm-svn: 268408
`Loop::makeLoopInvariant` can hoist instructions out of loops, so loop
dispositions for the loop it operated on may need to be cleared. We can
be smarter here (especially around how `forgetLoopDispositions` is
implemented), but let's be correct first.
Fixes PR27570.
llvm-svn: 268406
A few benchmarks with lots of accesses to global variables in the hot
loops regressed a lot since r266399, which added the
SpeculativeExecution pass to the default pipeline. The problem is that
this pass doesn't mark Globals Alias Analysis as preserved. Globals
Alias Analysis is computed in a module pass, whereas
SpeculativeExecution is a function pass, and a lot of passes dependent
on the Globals Alias Analysis to optimize these benchmarks are also
function passes. As such, the Globals Alias Analysis information cannot
be recomputed between SpeculativeExecution and the following function
passes needing that information.
SpeculativeExecution doesn't invalidate Globals Alias Analysis, so mark
it as such to fix those performance regressions.
Differential Revision: http://reviews.llvm.org/D19806
llvm-svn: 268370
We were overly cautious in our analysis of loops which have invokes
which unwind to EH pads. The loop unroll transform is safe because it
only clones blocks in the loop body, it does not try to split critical
edges involving EH pads. Instead, move the necessary safety check to
LoopUnswitch.
N.B. The safety check for loop unswitch is covered by an existing test
which fails without it.
llvm-svn: 268357
If a guard call being lowered by LowerGuardIntrinsics has the
`!make.implicit` metadata attached, then reattach the metadata to the
branch in the resulting expanded form of the intrinsic. This allows us
to implement null checks as guards and still get the benefit of implicit
null checks.
llvm-svn: 268148
support multiple induction variables
This patch enable loop reroll for the following case:
for(int i=0; i<N; i += 2) {
S += *a++;
S += *a++;
};
Differential Revision: http://reviews.llvm.org/D16550
llvm-svn: 268147
This moves some logic added to EarlyCSE in rL268120 into
`llvm::isInstructionTriviallyDead`. Adds a test case for DCE to
demonstrate that passes other than EarlyCSE can now pick up on the new
information.
llvm-svn: 268126
Summary:
This change teaches EarlyCSE some basic properties of guard intrinsics:
- Guard intrinsics read all memory, but don't write to any memory
- After a guard has executed, the condition it was guarding on can be
assumed to be true
- Guard intrinsics on a constant `true` are no-ops
Reviewers: reames, hfinkel
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D19578
llvm-svn: 268120
Summary:
Historically, we had a switch in the Makefiles for turning on "expensive
checks". This has never been ported to the cmake build, but the
(dead-ish) code is still around.
This will also make it easier to turn it on in buildbots.
Reviewers: chandlerc
Subscribers: jyknight, mzolotukhin, RKSimon, gberry, llvm-commits
Differential Revision: http://reviews.llvm.org/D19723
llvm-svn: 268050
I closely followed the precedents set by the vectorizer:
* With -Rpass-missed, the loop is reported with further details pointing
to -Rpass--analysis.
* -Rpass-analysis reports the details why distribution has failed.
* Regardless of -Rpass*, when distribution fails for a loop where
distribution was forced with the pragma, a warning is produced according
to -Wpass-failed. In this case the analysis info is also printed even
without -Rpass-analysis.
llvm-svn: 267952
The next patch will start using these for -Rpass-analysis so they won't
be internal-only anymore.
Move the 'Skipping; ' prefix that some of the message are using into the
'fail' function. We don't want to include this prefix in
the -Rpass-analysis report.
llvm-svn: 267951
"inferattrs" will deduce the attribute, but it will be too late for
many optimizations. Set it ourselves when creating the call.
Differential Revision: http://reviews.llvm.org/D17598
llvm-svn: 267762
This is required to use this function from isSafeToSpeculativelyExecute
Reviewed By: hfinkel
Differential Revision: http://reviews.llvm.org/D16231
llvm-svn: 267692
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
In the case where isLegalAddressingMode is used for cases
not related to addressing modes, such as pure adds and muls,
it should not be using address space 0. LSR already passes -1
as the address space in these cases.
llvm-svn: 267645
This splits out the per-loop functionality from the Pass class.
With this the fact whether the loop is forced-distribute with the new
metadata/pragma can be cached in the per-loop class rather than passed
around.
llvm-svn: 267643
We need the default ratio to be sufficiently large that it triggers transforms
based on block frequency info (BFI) and plays well with the recently introduced
BranchProbability used by CGP.
Differential Revision: http://reviews.llvm.org/D19435
llvm-svn: 267615
The original commit was reverted because of a buildbot problem with LazyCallGraph::SCC handling (not related to the OptBisect handling).
Differential Revision: http://reviews.llvm.org/D19172
llvm-svn: 267231
Summary: This change will shorten memset if the beginning of memset is overwritten by later stores.
Reviewers: hfinkel, eeckstein, dberlin, mcrosier
Subscribers: mgrang, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D18906
llvm-svn: 267197
We take the intersection of overflow flags while CSE'ing.
This permits us to consider two instructions with different overflow
behavior to be replaceable.
llvm-svn: 267153
EarlyCSE had inconsistent behavior with regards to flag'd instructions:
- In some cases, it would pessimize if the available instruction had
different flags by not performing CSE.
- In other cases, it would miscompile if it replaced an instruction
which had no flags with an instruction which has flags.
Fix this by being more consistent with our flag handling by utilizing
andIRFlags.
llvm-svn: 267111
This patch implements a optimization bisect feature, which will allow optimizations to be selectively disabled at compile time in order to track down test failures that are caused by incorrect optimizations.
The bisection is enabled using a new command line option (-opt-bisect-limit). Individual passes that may be skipped call the OptBisect object (via an LLVMContext) to see if they should be skipped based on the bisect limit. A finer level of control (disabling individual transformations) can be managed through an addition OptBisect method, but this is not yet used.
The skip checking in this implementation is based on (and replaces) the skipOptnoneFunction check. Where that check was being called, a new call has been inserted in its place which checks the bisect limit and the optnone attribute. A new function call has been added for module and SCC passes that behaves in a similar way.
Differential Revision: http://reviews.llvm.org/D19172
llvm-svn: 267022
"Into" was misleading. I am also planning to use this helper to look
for loop metadata and return the argument, so find seems like a better
name.
llvm-svn: 267013
This patch improves SimplifyCFG to catch cases like:
if (a < b) {
if (a > b) <- known to be false
unreachable;
}
Phabricator Revision: http://reviews.llvm.org/D18905
llvm-svn: 266767
Removed some unused headers, replaced some headers with forward class declarations.
Found using simple scripts like this one:
clear && ack --cpp -l '#include "llvm/ADT/IndexedMap.h"' | xargs grep -L 'IndexedMap[<]' | xargs grep -n --color=auto 'IndexedMap'
Patch by Eugene Kosov <claprix@yandex.ru>
Differential Revision: http://reviews.llvm.org/D19219
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 266595
Apparently there isn't test coverage for all of these. I'd appreciate
if someone with could reproduce and send me something to reduce, but for
now I've just looked for users of RemapInstruction and MapValue and
ensured they don't accidentally insert nullptr. Here is one of the
bootstraps that caught:
http://lab.llvm.org:8011/builders/clang-x64-ninja-win7/builds/11494
llvm-svn: 266567
Summary:
This lets us add this pass to the IR pass manager unconditionally; it
will simply not do anything on targets without branch divergence.
Reviewers: tra
Subscribers: llvm-commits, jingyue, rnk, chandlerc
Differential Revision: http://reviews.llvm.org/D18625
llvm-svn: 266398
Summary:
This fully solves the problem where the StructurizeCFG pass does not
consider the same branches as uniform as the SIAnnotateControlFlow pass.
The patch in D19013 helps with this problem, but is not sufficient
(and, interestingly, causes a "regression" with one of the existing
test cases).
No tests included here, because tests in D19013 already cover this.
Reviewers: arsenm, tstellarAMD
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D19018
llvm-svn: 266346
At some point, ARM stopped getting any benefit from ConstantHoisting because
the pass called a different variant of getIntImmCost. Reimplementing the
correct variant revealed some problems, however:
+ ConstantHoisting was modifying switch statements. This is simply invalid,
the cases must remain integer constants no matter the notional cost.
+ ConstantHoisting was mangling alloca instructions in the entry block. These
should be handled by FrameLowering, so constants actually have a cost of 0.
Worse, the resulting bitcasts meant they became dynamic allocas.
rdar://25707382
llvm-svn: 266260
LLVM optimization passes may reduce a profiled target expression
to a constant. Removing runtime calls at such instrumentation points
would help speedup the runtime of the instrumented program.
llvm-svn: 266229
Summary:
Fixes PR26774.
If you're aware of the issue, feel free to skip the "Motivation"
section and jump directly to "This patch".
Motivation:
I define "refinement" as discarding behaviors from a program that the
optimizer has license to discard. So transforming:
```
void f(unsigned x) {
unsigned t = 5 / x;
(void)t;
}
```
to
```
void f(unsigned x) { }
```
is refinement, since the behavior went from "if x == 0 then undefined
else nothing" to "nothing" (the optimizer has license to discard
undefined behavior).
Refinement is a fundamental aspect of many mid-level optimizations done
by LLVM. For instance, transforming `x == (x + 1)` to `false` also
involves refinement since the expression's value went from "if x is
`undef` then { `true` or `false` } else { `false` }" to "`false`" (by
definition, the optimizer has license to fold `undef` to any non-`undef`
value).
Unfortunately, refinement implies that the optimizer cannot assume
that the implementation of a function it can see has all of the
behavior an unoptimized or a differently optimized version of the same
function can have. This is a problem for functions with comdat
linkage, where a function can be replaced by an unoptimized or a
differently optimized version of the same source level function.
For instance, FunctionAttrs cannot assume a comdat function is
actually `readnone` even if it does not have any loads or stores in
it; since there may have been loads and stores in the "original
function" that were refined out in the currently visible variant, and
at the link step the linker may in fact choose an implementation with
a load or a store. As an example, consider a function that does two
atomic loads from the same memory location, and writes to memory only
if the two values are not equal. The optimizer is allowed to refine
this function by first CSE'ing the two loads, and the folding the
comparision to always report that the two values are equal. Such a
refined variant will look like it is `readonly`. However, the
unoptimized version of the function can still write to memory (since
the two loads //can// result in different values), and selecting the
unoptimized version at link time will retroactively invalidate
transforms we may have done under the assumption that the function
does not write to memory.
Note: this is not just a problem with atomics or with linking
differently optimized object files. See PR26774 for more realistic
examples that involved neither.
This patch:
This change introduces a new set of linkage types, predicated as
`GlobalValue::mayBeDerefined` that returns true if the linkage type
allows a function to be replaced by a differently optimized variant at
link time. It then changes a set of IPO passes to bail out if they see
such a function.
Reviewers: chandlerc, hfinkel, dexonsmith, joker.eph, rnk
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D18634
llvm-svn: 265762
As suggested by Chandler in his review comments for D18662, this
follow-on patch renames some variables in GetLoadValueForLoad and
CoerceAvailableValueToLoadType to hopefully make it more obvious
which variables hold value sizes and which hold load/store sizes.
No functional change intended.
llvm-svn: 265687
When GVN wants to re-interpret an already available value in a smaller
type, it needs to right-shift the value on big-endian systems to ensure
the correct bytes are accessed. The shift value is the difference of
the sizes of the two types.
This is correct as long as both types occupy multiples of full bytes.
However, when one of them is a sub-byte type like i1, this no longer
holds true: we still need to shift, but only to access the correct
*byte*. Accessing bits within the byte requires no shift in either
endianness; e.g. an i1 resides in the least-significant bit of its
containing byte on both big- and little-endian systems.
Therefore, the appropriate shift value to be used is the difference of
the *storage* sizes of the two types. This is already handled correctly
in one place where such a shift takes place (GetStoreValueForLoad), but
is incorrect in two other places: GetLoadValueForLoad and
CoerceAvailableValueToLoadType.
This patch changes both places to use the storage size as well.
Differential Revision: http://reviews.llvm.org/D18662
llvm-svn: 265684
Clarify what this RemapFlag actually means.
- Change the flag name to match its intended behaviour.
- Clearly document that it's not supposed to affect globals.
- Add a host of FIXMEs to indicate how to fix the behaviour to match
the intent of the flag.
RF_IgnoreMissingLocals should only affect the behaviour of
RemapInstruction for function-local operands; namely, for operands of
type Argument, Instruction, and BasicBlock. Currently, it is *only*
passed into RemapInstruction calls (and the transitive MapValue calls
that it makes).
When I split Metadata from Value I didn't understand the flag, and I
used it in a bunch of places for "global" metadata.
This commit doesn't have any functionality change, but prepares to
cleanup MapMetadata and MapValue.
llvm-svn: 265628
Summary:
In the context of http://wg21.link/lwg2445 C++ uses the concept of
'stronger' ordering but doesn't define it properly. This should be fixed
in C++17 barring a small question that's still open.
The code currently plays fast and loose with the AtomicOrdering
enum. Using an enum class is one step towards tightening things. I later
also want to tighten related enums, such as clang's
AtomicOrderingKind (which should be shared with LLVM as a 'C++ ABI'
enum).
This change touches a few lines of code which can be improved later, I'd
like to keep it as NFC for now as it's already quite complex. I have
related changes for clang.
As a follow-up I'll add:
bool operator<(AtomicOrdering, AtomicOrdering) = delete;
bool operator>(AtomicOrdering, AtomicOrdering) = delete;
bool operator<=(AtomicOrdering, AtomicOrdering) = delete;
bool operator>=(AtomicOrdering, AtomicOrdering) = delete;
This is separate so that clang and LLVM changes don't need to be in sync.
Reviewers: jyknight, reames
Subscribers: jyknight, llvm-commits
Differential Revision: http://reviews.llvm.org/D18775
llvm-svn: 265602
1. Add FullUnrollMaxCount option that works like MaxCount, but also limits
the unroll count for fully unrolled loops. So if a loop has an iteration
count over this, it won't fully unroll.
2. Add CLI options for MaxCount and the new option, so they can be tested
(plus a test).
3. Make partial unrolling obey MaxCount.
An example use-case (the out of tree one this is originally designed for) is
a target’s TTI can analyze a loop and decide on a max unroll count separate
from the size threshold, e.g. based on register pressure, then constrain
LoopUnroll to not exceed that, regardless of the size of the unrolled loop.
llvm-svn: 265562
Don't emit a gc.result for a statepoint lowered from
@llvm.experimental.deoptimize since the call into __llvm_deoptimize is
effectively noreturn. Instead follow the corresponding gc.statepoint
with an "unreachable".
llvm-svn: 265485
Summary:
As discussed on llvm-dev[1].
This change adds the basic boilerplate code around having this intrinsic
in LLVM:
- Changes in Intrinsics.td, and the IR Verifier
- A lowering pass to lower @llvm.experimental.guard to normal
control flow
- Inliner support
[1]: http://lists.llvm.org/pipermail/llvm-dev/2016-February/095523.html
Reviewers: reames, atrick, chandlerc, rnk, JosephTremoulet, echristo
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D18527
llvm-svn: 264976
Widening a PHI requires us to insert a trunc.
The logical place for this trunc is in the same BB as the PHI.
This is not possible if the BB is terminated by a catchswitch.
This fixes PR27133.
llvm-svn: 264926
This is effectively NFC, minus the renaming of the options
(-cyclone-prefetch-distance -> -prefetch-distance).
The change was requested by Tim in D17943.
llvm-svn: 264806
During ADCE, track which debug info scopes still have live references
from the code, and delete debug info intrinsics for the dead ones.
These intrinsics describe the locations of variables (in registers or
stack slots). If there's no code left corresponding to a variable's
scope, then there's no way to reference the variable in the debugger and
it doesn't matter what its value is.
I add a DEBUG printout when the described location in an SSA register,
in case it helps some trying to track down why locations get lost.
However, we still delete these; the scope itself isn't attached to any
real code, so the ship has already sailed.
llvm-svn: 264800
When eliminating or merging almost empty basic blocks, the existence of non-trivial PHI nodes
is currently used to recognize potential loops of which the block is the header and keep the block.
However, the current algorithm fails if the loops' exit condition is evaluated only with volatile
values hence no PHI nodes in the header. Especially when such a loop is an outer loop of a nested
loop, the loop is collapsed into a single loop which prevent later optimizations from being
applied (e.g., transforming nested loops into simplified forms and loop vectorization).
The patch augments the existing PHI node-based check by adding a pre-test if the BB actually
belongs to a set of loop headers and not eliminating it if yes.
llvm-svn: 264697
When eliminating or merging almost empty basic blocks, the existence of non-trivial PHI nodes
is currently used to recognize potential loops of which the block is the header and keep the block.
However, the current algorithm fails if the loops' exit condition is evaluated only with volatile
values hence no PHI nodes in the header. Especially when such a loop is an outer loop of a nested
loop, the loop is collapsed into a single loop which prevent later optimizations from being
applied (e.g., transforming nested loops into simplified forms and loop vectorization).
The patch augments the existing PHI node-based check by adding a pre-test if the BB actually
belongs to a set of loop headers and not eliminating it if yes.
llvm-svn: 264596
This changes RS4GC to lower calls to ``@llvm.experimental.deoptimize``
to gc.statepoints wrapping ``__llvm_deoptimize``, and changes
``callsGCLeafFunction`` to recognize ``@llvm.experimental.deoptimize``
as a non GC leaf function.
I've had to hard code the ``"__llvm_deoptimize"`` name in
RewriteStatepointsForGC; since ``TargetLibraryInfo`` is available only
during codegen. This isn't without precedent in the codebase, so I'm
not overtly concerned.
llvm-svn: 264456
We try to hoist the insertion point as high as possible to encourage
sharing. However, we must be careful not to hoist into a catchswitch as
it is both an EHPad and a terminator.
llvm-svn: 264344
isDependenceDistanceOfOne asserts that the store and the load access
through the same type. This function is also used by
removeDependencesFromMultipleStores so we need to make sure we filter
out mismatching types before reaching this point.
Now we do this when the initial candidates are gathered.
This is a refinement of the fix made in r262267.
Fixes PR27048.
llvm-svn: 264313
It's a bug fix.
For rerolled loops SE trip count remains unchanged. It leads to incorrect work of the next passes.
My patch just resets SE info for rerolled loop forcing SE to re-evaluate it next time it requested.
I also added a verifier call in the exisitng test to be sure no invalid SE data remain. Without my fix this test would fail with -verify-scev.
Differential Revision: http://reviews.llvm.org/D18316
llvm-svn: 264051
Summary:
It can hurt performance to prefetch ahead too much. Be conservative for
now and don't prefetch ahead more than 3 iterations on Cyclone.
Reviewers: hfinkel
Subscribers: llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D17949
llvm-svn: 263772
Summary:
And use this TTI for Cyclone. As it was explained in the original RFC
(http://thread.gmane.org/gmane.comp.compilers.llvm.devel/92758), the HW
prefetcher work up to 2KB strides.
I am also adding tests for this and the previous change (D17943):
* Cyclone prefetching accesses with a large stride
* Cyclone not prefetching accesses with a small stride
* Generic Aarch64 subtarget not prefetching either
Reviewers: hfinkel
Subscribers: aemerson, rengolin, llvm-commits, mzolotukhin
Differential Revision: http://reviews.llvm.org/D17945
llvm-svn: 263771
Summary:
If we decide to version a loop to benefit a transformation, it makes
sense to record the now non-aliasing accesses in the newly versioned
loop. This allows non-aliasing information to be used by subsequent
passes.
One example is 456.hmmer in SPECint2006 where after loop distribution,
we vectorize one of the newly distributed loops. To vectorize we
version this loop to fully disambiguate may-aliasing accesses. If we
add the noalias markers, we can use the same information in a later DSE
pass to eliminate some dead stores which amounts to ~25% of the
instructions of this hot memory-pipeline-bound loop. The overall
performance improves by 18% on our ARM64.
The scoped noalias annotation is added in LoopVersioning. The patch
then enables this for loop distribution. A follow-on patch will enable
it for the vectorizer. Eventually this should be run by default when
versioning the loop but first I'd like to get some feedback whether my
understanding and application of scoped noalias metadata is correct.
Essentially my approach was to have a separate alias domain for each
versioning of the loop. For example, if we first version in loop
distribution and then in vectorization of the distributed loops, we have
a different set of memchecks for each versioning. By keeping the scopes
in different domains they can conveniently be defined independently
since different alias domains don't affect each other.
As written, I also have a separate domain for each loop. This is not
necessary and we could save some metadata here by using the same domain
across the different loops. I don't think it's a big deal either way.
Probably the best is to review the tests first to see if I mapped this
problem correctly to scoped noalias markers. I have plenty of comments
in the tests.
Note that the interface is prepared for the vectorizer which needs the
annotateInstWithNoAlias API. The vectorizer does not use LoopVersioning
so we need a way to pass in the versioned instructions. This is also
why the maps have to become part of the object state.
Also currently, we only have an AA-aware DSE after the vectorizer if we
also run the LTO pipeline. Depending how widely this triggers we may
want to schedule a DSE toward the end of the regular pass pipeline.
Reviewers: hfinkel, nadav, ashutosh.nema
Subscribers: mssimpso, aemerson, llvm-commits, mcrosier
Differential Revision: http://reviews.llvm.org/D16712
llvm-svn: 263743
This splits out the logic that maps the `"statepoint-id"` attribute into
the actual statepoint ID, and the `"statepoint-num-patch-bytes"`
attribute into the number of patchable bytes the statpeoint is lowered
into. The new home of this logic is in IR/Statepoint.cpp, and this
refactoring will support similar functionality when lowering calls with
deopt operand bundles in the future.
llvm-svn: 263685
Summary:
Fix LSRInstance::HoistInsertPosition() to check the original insert
position block first for a canonical insertion point that is dominated
by all inputs. This leads to SCEV being able to reuse more instructions
since it currently tracks the instructions it creates for reuse by
keeping a table of <Value, insert point> pairs.
Reviewers: atrick
Subscribers: mcrosier, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D18001
llvm-svn: 263644
Adam Nemet