We were bailing to two places if our runtime checks failed. If the initial overflow check failed, we'd go to ScalarPH. If any other check failed, we'd go to MiddleBlock. This caused us to have to have an extra PHI per induction and reduction as the vector loop's exit block was not dominated by its latch.
There's no need to have this behavior - if we just always go to ScalarPH we can get rid of a bunch of complexity.
llvm-svn: 246637
This reduces the complexity of createEmptyBlock() and will open the door to further refactoring.
The test change is simply because we're now constant folding a trivial test.
llvm-svn: 246634
It makes things easier to understand if this is in a helper method. This is part of my ongoing spaghetti-removal operation on createEmptyLoop.
llvm-svn: 246632
There's no need to widen canonical induction variables. It's just as efficient to create a *new*, wide, induction variable.
Consider, if we widen an indvar, then we'll have to truncate it before its uses anyway (1 trunc). If we create a new indvar instead, we'll have to truncate that instead (1 trunc) [besides which IndVars should go and clean up our mess after us anyway on principle].
This lets us remove a ton of special-casing code.
llvm-svn: 246631
Vectorized loops only ever have one induction variable. All induction PHIs from the scalar loop are rewritten to be in terms of this single indvar.
We were trying very hard to pick an indvar that already existed, even if that indvar wasn't canonical (didn't start at zero). But trying so hard is really fruitless - creating a new, canonical, indvar only results in one extra add in the worst case and that add is trivially easy to push through the PHI out of the loop by instcombine.
If we try and be less clever here and instead let instcombine clean up our mess (as we do in many other places in LV), we can remove unneeded complexity.
llvm-svn: 246630
Teach FunctionAttr to infer the nonnull attribute on return values of functions which never return a potentially null value. This is done both via a conservative local analysis for the function itself and a optimistic per-SCC analysis. If no function in the SCC returns anything which could be null (other than values from other functions in the SCC), we can conclude no function returned a null pointer. Even if some function within the SCC returns a null pointer, we may be able to locally conclude that some don't.
Differential Revision: http://reviews.llvm.org/D9688
llvm-svn: 246476
Summary:
JumpThreading shouldn't duplicate a convergent call, because that would move a convergent call into a control-inequivalent location. For example,
if (cond) {
...
} else {
...
}
convergent_call();
if (cond) {
...
} else {
...
}
should not be optimized to
if (cond) {
...
convergent_call();
...
} else {
...
convergent_call();
...
}
Test Plan: test/Transforms/JumpThreading/basic.ll
Patch by Xuetian Weng.
Reviewers: resistor, arsenm, jingyue
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D12484
llvm-svn: 246415
PR24605 is caused due to an incorrect insert point in instcombine's IR
builder. When simplifying
%t = add X Y
...
%m = icmp ... %t
the replacement for %t should be placed before %t, not before %m, as
there could be a use of %t between %t and %m.
llvm-svn: 246315
Summary:
This patch removes two remaining places where pointer value comparisons
are used to order functions: comparing range annotation metadata, and comparing
block address constants. (These are both rare cases, and so no actual
non-determinism was observed from either case).
The fix for range metadata is simple: the annotation always consists of a pair
of integers, so we just order by those integers.
The fix for block addresses is more subtle. Two constants are the same if they
are the same basic block in the same function, or if they refer to corresponding
basic blocks in each respective function. Note that in the first case, merging
is trivially correct. In the second, the correctness of merging relies on the
fact that the the values of block addresses cannot be compared. This change is
actually an enhancement, as these functions could not previously be merged (see
merge-block-address.ll).
There is still a problem with cross function block addresses, in that constants
pointing to a basic block in a merged function is not updated.
This also more robustly compares floating point constants by all fields of their
semantics, and fixes a dyn_cast/cast mixup.
Author: jrkoenig
Reviewers: dschuff, nlewycky, jfb
Subscribers llvm-commits
Differential revision: http://reviews.llvm.org/D12376
llvm-svn: 246305
handle more allocas with loads past the end of the alloca.
I suspect there are some related crashers with slightly different
patterns, but I'll fix those and add test cases as I find them.
Thanks to David Majnemer for the excellent test case reduction here.
Made this super simple to debug and fix.
llvm-svn: 246289
After hitting @llvm.assume(X) we can:
- propagate equality that X == true
- if X is icmp/fcmp (with eq operation), and one of operand
is constant we can change all variables with constants in the same BasicBlock
http://reviews.llvm.org/D11918
llvm-svn: 246243
This patch changes the analysis diagnostics produced when loops with
floating-point recurrences or memory operations are identified. The new messages
say "cannot prove it is safe to reorder * operations; allow reordering by
specifying #pragma clang loop vectorize(enable)". Depending on the type of
diagnostic the message will include additional options such as ffast-math or
__restrict__.
This patch also allows the vectorize(enable) pragma to override the low pointer
memory check threshold. When the hint is given a higher threshold is used.
See the clang patch for the options produced for each diagnostic.
llvm-svn: 246187
Unlike scalar operations, we can perform vector operations on element types that
are smaller than the native integer types. We type-promote scalar operations if
they are smaller than a native type (e.g., i8 arithmetic is promoted to i32
arithmetic on Arm targets). This patch detects and removes type-promotions
within the reduction detection framework, enabling the vectorization of small
size reductions.
In the legality phase, we look through the ANDs and extensions that InstCombine
creates during promotion, keeping track of the smaller type. In the
profitability phase, we use the smaller type and ignore the ANDs and extensions
in the cost model. Finally, in the code generation phase, we truncate the result
of the reduction to allow InstCombine to rewrite the entire expression in the
smaller type.
This fixes PR21369.
http://reviews.llvm.org/D12202
Patch by Matt Simpson <mssimpso@codeaurora.org>!
llvm-svn: 246149
... and move it into LoopUtils where it can be used by other passes, just like ReductionDescriptor. The API is very similar to ReductionDescriptor - that is, not very nice at all. Sorting these both out will come in a followup.
NFC
llvm-svn: 246145
A release fence acts as a publication barrier for stores within the current thread to become visible to other threads which might observe the release fence. It does not require the current thread to observe stores performed on other threads. As a result, we can allow store-load and load-store forwarding across a release fence.
We do need to make sure that stores before the fence can't be eliminated even if there's another store to the same location after the fence. In theory, we could reorder the second store above the fence and *then* eliminate the former, but we can't do this if the stores are on opposite sides of the fence.
Note: While more aggressive then what's there, this patch is still implementing a really conservative ordering. In particular, I'm not trying to exploit undefined behavior via races, or the fact that the LangRef says only 'atomic' accesses are ordered w.r.t. fences.
Differential Revision: http://reviews.llvm.org/D11434
llvm-svn: 246134
When computing base pointers, we introduce new instructions to propagate the base of existing instructions which might not be bases. However, the algorithm doesn't make any effort to recognize when the new instruction to be inserted is the same as an existing one already in the IR. Since this is happening immediately before rewriting, we don't really have a chance to fix it after the pass runs without teaching loop passes about statepoints.
I'm really not thrilled with this patch. I've rewritten it 4 different ways now, but this is the best I've come up with. The case where the new instruction is just the original base defining value could be merged into the existing algorithm with some complexity. The problem is that we might have something like an extractelement from a phi of two vectors. It may be trivially obvious that the base of the 0th element is an existing instruction, but I can't see how to make the algorithm itself figure that out. Thus, I resort to the call to SimplifyInstruction instead.
Note that we can only adjust the instructions we've inserted ourselves. The live sets are still being tracked in side structures at this point in the code. We can't easily muck with instructions which might be in them. Long term, I'm really thinking we need to materialize the live pointer sets explicitly in the IR somehow rather than using side structures to track them.
Differential Revision: http://reviews.llvm.org/D12004
llvm-svn: 246133
This patch ensures that every analysis diagnostic produced by the vectorizer
will be printed if the loop has a vectorization hint on it. The condition has
also been improved to prevent printing when a disabling hint is specified.
llvm-svn: 246132
As Sanjoy pointed out over in http://reviews.llvm.org/D11819, a switch on an icmp should always be able to become a branch instruction. This patch generalizes that notion slightly to prove that the default case of a switch is unreachable if the cases completely cover all possible bit patterns in the condition. Once that's done, the switch to branch conversion kicks in just fine.
Note: Duplicate case values are disallowed by the LangRef and verifier.
Differential Revision: http://reviews.llvm.org/D11995
llvm-svn: 246125
The problem here were the function analyses invoked by the function pass
manager from the new IPO pass. I looked at other IPO passes needing
dominance information and the only one that requires it (partial
inliner) does not use the standard dependency mechanism.
This patch mimics what the partial inliner does to compute dominance,
post-dominance and loop info. One thing I like about this approach is
that I can delay the computation of all this until I actually need it.
This should bring the ASAN buildbot back to green. If there's a better
way to fix this, I'll do it in a follow-up patch.
llvm-svn: 246066
This was only added to preserve the old ScalarRepl's use of SSAUpdater
which was originally to avoid use of dominance frontiers. Now, we only
need a domtree, and we'll need a domtree right after this pass as well
and so it makes perfect sense to always and only use the dom-tree
powered mem2reg. This was flag-flipper earlier and has stuck reasonably
so I wanted to gut the now-dead code out of SROA before we waste more
time with it. Among other things, this will make passmanager porting
easier.
llvm-svn: 246028
Summary: When comparing basic blocks, there is an additional check that two Value*'s should have the same ID, which interferes with merging equivalent constants of different kinds (such as a ConstantInt and a ConstantPointerNull in the included testcase). The cmpValues function already ensures that the two values in each function are the same, so removing this check should not cause incorrect merging.
Also, the type comparison is redundant, based on reviewing the code and testing on the test suite and several large LTO bitcodes.
Author: jrkoenig
Reviewers: nlewycky, jfb, dschuff
Subscribers: llvm-commits
Differential revision: http://reviews.llvm.org/D12302
llvm-svn: 246001
Summary:
This change makes the variable argument intrinsics, `llvm.va_start` and
`llvm.va_copy`, and the `va_arg` instruction behave as they do on Windows
inside a `CallingConv::X86_64_Win64` function. It's needed for a Clang patch
I have to add support for GCC's `__builtin_ms_va_list` constructs.
Reviewers: nadav, asl, eugenis
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D1622
llvm-svn: 245990
Extend signed relational comparison instrumentation with a special
case for comparisons with -1. This fixes an MSan false positive when
such comparison is used as a sign bit test.
https://llvm.org/bugs/show_bug.cgi?id=24561
llvm-svn: 245980
The loop minimum iterations check below ensures the loop has enough trip count so the generated
vector loop will likely be executed, and it covers the overflow check.
Differential Revision: http://reviews.llvm.org/D12107.
llvm-svn: 245952
Eventually, we will need sample profiles to be incorporated into the
inliner's cost models. To do this, we need the sample profile pass to
be a module pass.
This patch makes no functional changes beyond the mechanical adjustments
needed to run SampleProfile as a module pass.
llvm-svn: 245940
It doesn't solve the problem, when for example we load something, and
then assume that it is the same as some constant value, because
globalopt will fail on unknown load instruction. The proposed solution
would be to skip some instructions that we can't evaluate and they are
safe to skip (f.e. load, assume and many others) and see if they are
required to perform optimization (f.e. we don't care about ephemeral
instructions that may appear using @llvm.assume())
http://reviews.llvm.org/D12266
llvm-svn: 245919
This patch adds support for dfsan on aarch64-linux with 42-bit VMA
(current default config for 64K pagesize kernels). The support is
enabled by defining the SANITIZER_AARCH64_VMA to 42 at build time
for both clang/llvm and compiler-rt. The default VMA is 39 bits.
llvm-svn: 245840
TL-DR: SROA is followed by EarlyCSE which requires the DominatorTree.
There is no reason not to require it up-front for SROA.
Some history is necessary to understand why we ended-up here.
r123437 switched the second (Legacy)SROA in the optimizer pipeline to
use SSAUpdater in order to avoid recomputing the costly
DominanceFrontier. The purpose was to speed-up the compile-time.
Later r123609 removed the need for the DominanceFrontier in
(Legacy)SROA.
Right after, some cleanup was made in r123724 to remove any reference
to the DominanceFrontier. SROA existed in two flavors: SROA_SSAUp and
SROA_DT (the latter replacing SROA_DF).
The second argument of `createScalarReplAggregatesPass` was renamed
from `UseDomFrontier` to `UseDomTree`.
I believe this is were a mistake was made. The pipeline was not
updated and the call site was still:
PM->add(createScalarReplAggregatesPass(-1, false));
At that time, SROA was immediately followed in the pipeline by
EarlyCSE which required alread the DominatorTree. Not requiring
the DominatorTree in SROA didn't save anything, but unfortunately
it was lost at this point.
When the new SROA Pass was introduced in r163965, I believe the goal
was to have an exact replacement of the existing SROA, this bug
slipped through.
You can see currently:
$ echo "" | clang -x c++ -O3 -c - -mllvm -debug-pass=Structure
...
...
FunctionPass Manager
SROA
Dominator Tree Construction
Early CSE
After this patch:
$ echo "" | clang -x c++ -O3 -c - -mllvm -debug-pass=Structure
...
...
FunctionPass Manager
Dominator Tree Construction
SROA
Early CSE
This improves the compile time from 88s to 23s for PR17855.
https://llvm.org/bugs/show_bug.cgi?id=17855
And from 113s to 12s for PR16756
https://llvm.org/bugs/show_bug.cgi?id=16756
Reviewers: chandlerc
Differential Revision: http://reviews.llvm.org/D12267
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 245820
Summary:
WinEHPrepare is going to require that cleanuppad and catchpad produce values
of token type which are consumed by any cleanupret or catchret exiting the
pad. This change updates the signatures of those operators to require/enforce
that the type produced by the pads is token type and that the rets have an
appropriate argument.
The catchpad argument of a `CatchReturnInst` must be a `CatchPadInst` (and
similarly for `CleanupReturnInst`/`CleanupPadInst`). To accommodate that
restriction, this change adds a notion of an operator constraint to both
LLParser and BitcodeReader, allowing appropriate sentinels to be constructed
for forward references and appropriate error messages to be emitted for
illegal inputs.
Also add a verifier rule (noted in LangRef) that a catchpad with a catchpad
predecessor must have no other predecessors; this ensures that WinEHPrepare
will see the expected linear relationship between sibling catches on the
same try.
Lastly, remove some superfluous/vestigial casts from instruction operand
setters operating on BasicBlocks.
Reviewers: rnk, majnemer
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D12108
llvm-svn: 245797
Summary:
Merge functions previously relied on unsigned comparisons of pointer values to
order functions. This caused observable non-determinism in the compiler for
large bitcode programs. Basically, opt -mergefuncs program.bc | md5sum produces
different hashes when run repeatedly on the same machine. Differing output was
observed on three large bitcodes, but it was less frequent on the smallest file.
It is possible that this only manifests on the large inputs, hence remaining
undetected until now.
This patch fixes this by removing (almost, see below) all places where
comparisons between pointers are used to order functions. Most of these changes
are local, but the comparison of global values requires assigning an identifier
to each local in the order it is visited. This is very similar to the way the
comparison function identifies Value*'s defined within a function. Because the
order of visiting the functions and their subparts is deterministic, the
identifiers assigned to the globals will be as well, and the order of functions
will be deterministic.
With these changes, there is no more observed non-determinism. There is also
only minor slowdowns (negligible to 4%) compared to the baseline, which is
likely a result of the fact that global comparisons involve hash lookups and not
just pointer comparisons.
The one caveat so far is that programs containing BlockAddress constants can
still be non-deterministic. It is not clear what the right solution is here. In
particular, even if the global numbers are used to order by function, we still
need a way to order the BasicBlock*'s. Unfortunately, we cannot just bail out
and fail to order the functions or consider them equal, because we require a
total order over functions. Note that programs with BlockAddress constants are
relatively rare, so the impact of leaving this in is minor as long as this pass
is opt-in.
Author: jrkoenig
Reviewers: nlewycky, jfb, dschuff
Subscribers: jevinskie, llvm-commits, chapuni
Differential revision: http://reviews.llvm.org/D12168
llvm-svn: 245762
The original checkin was buggy, this change has a fix.
Original commit message:
[InstCombine] Transform A & (L - 1) u< L --> L != 0
Summary:
This transform is never a pessimization at the IR level (since it
replaces an `icmp` with another), and has potentiall payoffs:
1. It may make the `icmp` fold away or become loop invariant.
2. It may make the `A & (L - 1)` computation dead.
This shows up in Java, in range checks generated by array accesses of
the form `a[i & (a.length - 1)]`.
Reviewers: reames, majnemer
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D12210
llvm-svn: 245753
Gets a bit tricky in the ValueMapper, of course - not sure if we should
just expose a list of explicit types for each Value so that the
ValueMapper can be neutral to these special cases (it's OK for things
like load, where the explicit type is the result type - but when that's
not the case, it means plumbing through another "special" type... )
llvm-svn: 245728
James Molloy