| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
|
|
|
|
| |
The memory location that corresponds to a volatile operation is very
special. They are observed by the machine in ways which we cannot
reason about.
Differential Revision: http://reviews.llvm.org/D20555
llvm-svn: 270879
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
It turns out that too many passes are relying on alias analysis results
for control dependencies. Until we fix that by introducing a more accurate
modelling of control dependencies, special case assume in MemorySSA instead.
Also introduce tests to ensure we don't regress the FunctionAttrs or LICM
passes.
Differential Revision: http://reviews.llvm.org/D20658
llvm-svn: 270823
|
| |
|
|
| |
llvm-svn: 270779
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If a we have (a) a GEP and (b) a pointer based on an alloca, and the
beginning of the object the GEP points would have a negative offset with
repsect to the alloca, then the GEP can not alias pointer (b).
For example, consider code like:
struct { int f0, int f1, ...} foo;
...
foo alloca;
foo *random = bar(alloca);
int *f0 = &alloca.f0
int *f1 = &random->f1;
Which is lowered, approximately, to:
%alloca = alloca %struct.foo
%random = call %struct.foo* @random(%struct.foo* %alloca)
%f0 = getelementptr inbounds %struct, %struct.foo* %alloca, i32 0, i32 0
%f1 = getelementptr inbounds %struct, %struct.foo* %random, i32 0, i32 1
Assume %f1 and %f0 alias. Then %f1 would point into the object allocated
by %alloca. Since the %f1 GEP is inbounds, that means %random must also
point into the same object. But since %f0 points to the beginning of %alloca,
the highest %f1 can be is (%alloca + 3). This means %random can not be higher
than (%alloca - 1), and so is not inbounds, a contradiction.
Differential Revision: http://reviews.llvm.org/D20495
llvm-svn: 270777
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Getting accurate locations for loops is important, because those locations are
used by the frontend to generate optimization remarks. Currently, optimization
remarks for loops often appear on the wrong line, often the first line of the
loop body instead of the loop itself. This is confusing because that line might
itself be another loop, or might be somewhere else completely if the body was
inlined function call. This happens because of the way we find the loop's
starting location. First, we look for a preheader, and if we find one, and its
terminator has a debug location, then we use that. Otherwise, we look for a
location on an instruction in the loop header.
The fallback heuristic is not bad, but will almost always find the beginning of
the body, and not the loop statement itself. The preheader location search
often fails because there's often not a preheader, and even when there is a
preheader, depending on how it was formed, it sometimes carries the location of
some preceeding code.
I don't see any good theoretical way to fix this problem. On the other hand,
this seems like a straightforward solution: Put the debug location in the
loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert
llvm.loop metadata with appropriate locations when generating debugging
information. With these changes, our loop remarks have much more accurate
locations.
Differential Revision: http://reviews.llvm.org/D19738
llvm-svn: 270771
|
| |
|
|
|
|
| |
My script missed those in r270750.
llvm-svn: 270763
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
There was a typo in r267758. It caused invalid accesses when
given something like "void @free(...)", as NumParams == 0, and
we then try to look at the 0th parameter.
Turns out, most of these were untested; add both attribute
and missing-prototype checks for all libc libfuncs.
Differential Revision: http://reviews.llvm.org/D20543
llvm-svn: 270750
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Summary:
**Description**
This makes `WidenIV::widenIVUse` (IndVarSimplify.cpp) fail to widen narrow IV uses in some cases. The latter affects IndVarSimplify which may not eliminate narrow IV's when there actually exists such a possibility, thereby producing ineffective code.
When `WidenIV::widenIVUse` gets a NarrowUse such as `{(-2 + %inc.lcssa),+,1}<nsw><%for.body3>`, it first tries to get a wide recurrence for it via the `getWideRecurrence` call.
`getWideRecurrence` returns recurrence like this: `{(sext i32 (-2 + %inc.lcssa) to i64),+,1}<nsw><%for.body3>`.
Then a wide use operation is generated by `cloneIVUser`. The generated wide use is evaluated to `{(-2 + (sext i32 %inc.lcssa to i64))<nsw>,+,1}<nsw><%for.body3>`, which is different from the `getWideRecurrence` result. `cloneIVUser` sees the difference and returns nullptr.
This patch also fixes the broken LLVM tests by adding missing <nsw> entries introduced by the correction.
**Minimal reproducer:**
```
int foo(int a, int b, int c);
int baz();
void bar()
{
int arr[20];
int i = 0;
for (i = 0; i < 4; ++i)
arr[i] = baz();
for (; i < 20; ++i)
arr[i] = foo(arr[i - 4], arr[i - 3], arr[i - 2]);
}
```
**Clang command line:**
```
clang++ -mllvm -debug -S -emit-llvm -O3 --target=aarch64-linux-elf test.cpp -o test.ir
```
**Expected result:**
The ` -mllvm -debug` log shows that all the IV's for the second `for` loop have been eliminated.
Reviewers: sanjoy
Subscribers: atrick, asl, aemerson, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D20058
llvm-svn: 270695
|
| |
|
|
|
|
| |
This fixes PR27847. Now for real.
llvm-svn: 270629
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Similar in spirit to D20497 :
If all elements of a constant vector are known non-zero, then we can say that the
whole vector is known non-zero.
It seems like we could extend this to FP scalar/vector too, but isKnownNonZero()
says it only works for integers and pointers for now.
Differential Revision: http://reviews.llvm.org/D20544
llvm-svn: 270562
|
| |
|
|
|
|
| |
This fixes PR27847.
llvm-svn: 270517
|
| |
|
|
| |
llvm-svn: 270465
|
| |
|
|
| |
llvm-svn: 270370
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
vector splat constants
We could try harder to handle non-splat vector constants too,
but that seems much rarer to me.
Note that the div test isn't resolved because there's a check
for isIntegerTy() guarding that transform.
Differential Revision: http://reviews.llvm.org/D20497
llvm-svn: 270369
|
| |
|
|
| |
llvm-svn: 270272
|
| |
|
|
|
|
|
|
| |
When it has a DataLayout, DecomposeGEPExpression() should return the same object
as GetUnderlyingObject(). Per the FIXME, it currently always has a DL, so the
runtime check is redundant and can become an assert.
llvm-svn: 270268
|
| |
|
|
|
|
|
|
| |
Before r257832, the threshold used by SimpleInliner was explicitly specified or generated from opt levels and passed to the base class Inliner's constructor. There, it was first overridden by explicitly specified -inline-threshold. The refactoring in r257832 did not preserve this behavior for all opt levels. This change brings back the original behavior.
Differential Revision: http://reviews.llvm.org/D20452
llvm-svn: 270153
|
| |
|
|
|
|
|
|
|
|
|
|
| |
This patch changes the order in which we attempt to prove the independence of
strided accesses. We previously did this after we knew the dependence distance
was positive. With this change, we check for independence before handling the
negative distance case. The patch prevents LAA from reporting forward
dependences for independent strided accesses.
This change was requested in the review of D19984.
llvm-svn: 270072
|
| |
|
|
|
|
|
|
|
|
| |
... for AddRec's in loops for which SCEV is unable to compute a max
tripcount. This is the NUW variant of r269211 and fixes PR27691.
(Note: PR27691 is not a correct or stability bug, it was created to
track a pending task).
llvm-svn: 269790
|
| |
|
|
|
|
|
| |
Original change Hal's comments were based on:
http://reviews.llvm.org/D19730
llvm-svn: 269678
|
| |
|
|
|
|
|
| |
This patch renames the option enabling the store-to-load forwarding conflict
detection optimization. This change was requested in the review of D20241.
llvm-svn: 269668
|
| |
|
|
|
|
|
| |
Also s/Cycles/Iters/ in NumCyclesForStoreLoadThroughMemory to make it
clear that this is not about clock cycles but loop cycles/iterations.
llvm-svn: 269667
|
| |
|
|
| |
llvm-svn: 269666
|
| |
|
|
| |
llvm-svn: 269654
|
| |
|
|
| |
llvm-svn: 269508
|
| |
|
|
| |
llvm-svn: 269507
|
| |
|
|
|
|
|
|
|
|
| |
increasing cost tracking system during the full unroll heuristic analysis that avoids counting any instruction cost until that instruction becomes "live" through a side-effect or use outside the...""
This reverts commit r269395.
Try to reapply with a fix from chapuni.
llvm-svn: 269486
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Fix "Logic error" warnings of the type "Called C++ object pointer is
null" reported by Clang Static Analyzer on the following files:
lib/Analysis/ScalarEvolution.cpp,
lib/Analysis/LoopInfo.cpp.
Patch by Apelete Seketeli!
llvm-svn: 269424
|
| |
|
|
|
|
|
|
|
|
|
| |
tracking system during the full unroll heuristic analysis that avoids counting any instruction cost until that instruction becomes "live" through a side-effect or use outside the..."
This reverts commit r269388.
It caused some bots to fail, I'm reverting it until I investigate the
issue.
llvm-svn: 269395
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
system during the full unroll heuristic analysis that avoids counting any instruction cost until that instruction becomes "live" through a side-effect or use outside the...
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
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
simplified.
Summary:
Currently we consider such instructions as simplified, which is incorrect,
because if their user isn't simplified, we can't actually simplify them too.
This biases our estimates of profitability: for instance the analyzer expects
much more gains from unrolling memcpy loops than there actually are.
Reviewers: hfinkel, chandlerc
Subscribers: mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D17365
llvm-svn: 269387
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
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
|
| |
|
|
| |
llvm-svn: 269356
|
| |
|
|
|
|
|
|
|
| |
SCEVExpander::replaceCongruentIVs assumes the backedge value of an
SCEV-analysable PHI to always be an instruction, when this is not
necessarily true. For now address this by bailing out of the
optimization if the backedge value of the PHI is a non-Instruction.
llvm-svn: 269213
|
| |
|
|
|
|
|
|
|
|
|
|
| |
`SCEVExpander::replaceCongruentIVs` bypasses `hoistIVInc` if both the
original and the isomorphic increments are PHI nodes. Doing this can
break SSA if the isomorphic increment is not dominated by the original
increment. Get rid of the bypass, and let `hoistIVInc` do the right
thing.
Fixes PR27232 (compile time crash/hang).
llvm-svn: 269212
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
... for AddRec's in loops for which SCEV is unable to compute a max
tripcount. This is not a problem for "normal" loops[0] that don't have
guards or assumes, but helps in cases where we have guards or assumes in
the loop that can be used to constrain incoming values over the backedge.
This partially fixes PR27691 (we still don't handle the NUW case).
[0]: for "normal" loops, in the cases where we'd be able to prove
no-wrap via isKnownPredicate, we'd also be able to compute a max
tripcount.
llvm-svn: 269211
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Equivalent GEP indices with different types are treated as different
indices altogether, leading to an incorrect AA result. Fix the issue
by comparing indices based on their values.
Thanks to Mikael Holmén for reporting the issue!
Differential Revision: http://reviews.llvm.org/D19935
llvm-svn: 269197
|
| |
|
|
|
|
|
|
|
|
| |
Extract a part of isDereferenceableAndAlignedPointer functionality to Value:
Reviewed By: hfinkel, sanjoy
Differential Revision: http://reviews.llvm.org/D17611
llvm-svn: 269190
|
| |
|
|
| |
llvm-svn: 269138
|
| |
|
|
| |
llvm-svn: 269131
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Do simplifications common to all shift instructions based on the amount shifted:
1. If the shift amount is known larger than the bitwidth, the result is undefined.
2. If the valid bits of the shift amount are all known to be 0, it's a shift by zero, so the shift operand is the result.
Note that we could generalize the shift-by-zero transform into a shift-by-constant if all of the valid bits in the shift
amount are known, but that would have to be done in InstCombine rather than here because it would mean we need to create
a new shift instruction.
Differential Revision: http://reviews.llvm.org/D19874
llvm-svn: 269114
|
| |
|
|
| |
llvm-svn: 269094
|
| |
|
|
|
|
| |
MSVC build.
llvm-svn: 269091
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Analysis.
The plan is to eventually make this logic simpler, however I expect it to
be a little tricky for the foreseeable future (at least until we're rid of
pointee types), so move it here so that it can be reused to build a summary
index for devirtualization.
Differential Revision: http://reviews.llvm.org/D20005
llvm-svn: 269081
|
| |
|
|
|
|
|
|
|
|
|
|
| |
This removes a redundant stride versioning step (we already
do it in getPtrStride, so it has no effect) and uses PSE to
get the SCEV expressions for the source and destination
(this might have changed when getPtrStride was called).
I discovered this through code inspection, and couldn't
produce a regression test for it.
llvm-svn: 269052
|
| |
|
|
|
|
|
|
| |
This was a fairly simple patch but on closer inspection was seriously flawed and caused PR27690.
This reverts commit r268921.
llvm-svn: 269051
|
| |
|
|
|
|
|
| |
Changing misleading function name was approved in http://reviews.llvm.org/D17268.
Patch by Roman Shirokiy.
llvm-svn: 269021
|
| |
|
|
|
|
|
|
|
|
| |
Reviewers: apilipenko, majnemer, reames
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D20044
llvm-svn: 269008
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Summary:
The idea is very close to what we do for assume intrinsics: we mark the
guard intrinsics as writing to arbitrary memory to maintain control
dependence, but under the covers we teach AA that they do not mod any
particular memory location.
Reviewers: chandlerc, hfinkel, gbiv, reames
Subscribers: george.burgess.iv, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D19575
llvm-svn: 269007
|
| |
|
|
|
|
| |
The boolean expressions are somewhat hard to read otherwise.
llvm-svn: 268998
|