The patch rL298481 was reverted due to crash on clang-with-lto-ubuntu build.
The reason of the crash was type mismatch between either a or b and RHS in the following situation:
LHS = sext(a +nsw b) > RHS.
This is quite rare, but still possible situation. Normally we need to cast all {a, b, RHS} to their widest type.
But we try to avoid creation of new SCEV that are not constants to avoid initiating recursive analysis that
can take a lot of time and/or cache a bad value for iterations number. To deal with this, in this patch we
reject this case and will not try to analyze it if the type of sum doesn't match with the type of RHS. In this
situation we don't need to create any non-constant SCEVs.
This patch also adds an assertion to the method IsProvedViaContext so that we could fail on it and not
go further into range analysis etc (because in some situations these analyzes succeed even when the passed
arguments have wrong types, what should not normally happen).
The patch also contains a fix for a problem with too narrow scope of the analysis caused by wrong
usage of predicates in recursive invocations.
The regression test on the said failure: test/Analysis/ScalarEvolution/implied-via-addition.ll
llvm-svn: 298690
Given below case:
%y = shl %x, n
%z = ashr %y, m
when n = m, SCEV models it as sext(trunc(x)). This patch tries to handle
the case where n > m by using sext(mul(trunc(x), 2^(n-m)))) as the SCEV
expression.
llvm-svn: 298631
Summary:
Adding a printer pass for printing the LVI cache values after transformations
that use LVI.
This will help us in identifying cases where LVI
invariants are violated, or transforms that leave LVI in an incorrect state.
Right now, I have added two test cases to show that the printer pass is working.
I will be adding more test cases in a later change, once this change is
checked in upstream.
Reviewers: reames, dberlin, sanjoy, apilipenko
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30790
llvm-svn: 298542
This patch allows SCEV predicate analysis to prove implication of some expression predicates
from context predicates related to arguments of those expressions.
It introduces three new rules:
For addition:
(A >X && B >= 0) || (B >= 0 && A > X) ===> (A + B) > X.
For division:
(A > X) && (0 < B <= X + 1) ===> (A / B > 0).
(A > X) && (-B <= X < 0) ===> (A / B >= 0).
Using these rules, SCEV is able to prove facts like "if X > 1 then X / 2 > 0".
They can also be combined with the same context, to prove more complex expressions like
"if X > 1 then X/2 + 1 > 1".
Diffirential Revision: https://reviews.llvm.org/D30887
Reviewed by: sanjoy
llvm-svn: 298481
Currently the default C calling convention functions are treated
the same as compute kernels. Make this explicit so the default
calling convention can be changed to a non-kernel.
Converted with perl -pi -e 's/define void/define amdgpu_kernel void/'
on the relevant test directories (and undoing in one place that actually
wanted a non-kernel).
llvm-svn: 298444
After the loop unroll threshold was increased in r295538, very
large constant expressions can be created. This prevents them
from having to be recursively scanned, leading to a compile
time blow-up.
Differential Revision: https://reviews.llvm.org/D30689
llvm-svn: 298356
If loop bound containing calculations like min(a,b), the Scalar
Evolution API getSmallConstantTripMultiple returns 4294967295 "-1"
as the trip multiple. The problem is that, SCEV use -1 * umax to
represent umin. The multiple constant -1 was returned, and the logic
of guarding against huge trip counts was skipped. Because -1 has 32
active bits.
The fix attempt to factor more general cases. First try to get the
greatest power of two divisor of trip count expression. In case
overflow happens, the trip count expression is still divisible by the
greatest power of two divisor returned. Returns 1 if not divisible by 2.
Patch by Huihui Zhang <huihuiz@codeaurora.org>
Differential Revision: https://reviews.llvm.org/D30840
llvm-svn: 298301
Summary:
This approach has two major advantages over the existing one:
1. We don't need to extend bitwidth in our computations. Extending
bitwidth is a big issue for compile time as we often end up working with
APInts wider than 64bit, which is a slow case for APInt.
2. When we zero extend a wrapped range, we lose some information (we
replace the range with [0, 1 << src bit width)). Thus, avoiding such
extensions better preserves information.
Correctness testing:
I ran 'ninja check' with assertions that the new implementation of
getRangeForAffineAR gives the same results as the old one (this
functionality is not present in this patch). There were several failures
- I inspected them manually and found out that they all are caused by
the fact that we're returning more accurate results now (see bullet (2)
above).
Without such assertions 'ninja check' works just fine, as well as
SPEC2006.
Compile time testing:
CTMark/Os:
- mafft/pairlocalalign -16.98%
- tramp3d-v4/tramp3d-v4 -12.72%
- lencod/lencod -11.51%
- Bullet/bullet -4.36%
- ClamAV/clamscan -3.66%
- 7zip/7zip-benchmark -3.19%
- sqlite3/sqlite3 -2.95%
- SPASS/SPASS -2.74%
- Average -5.81%
Performance testing:
The changes are expected to be neutral for runtime performance.
Reviewers: sanjoy, atrick, pete
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30477
llvm-svn: 297992
getIntrinsicInstrCost() used to only compute scalarization cost based on types.
This patch improves this so that the actual arguments are checked when they are
available, in order to handle only unique non-constant operands.
Tests updates:
Analysis/CostModel/X86/arith-fp.ll
Transforms/LoopVectorize/AArch64/interleaved_cost.ll
Transforms/LoopVectorize/ARM/interleaved_cost.ll
The improvement in getOperandsScalarizationOverhead() to differentiate on
constants made it necessary to update the interleaved_cost.ll tests even
though they do not relate to intrinsics.
Review: Hal Finkel
https://reviews.llvm.org/D29540
llvm-svn: 297705
When the array indexes are all determined by GVN to be constants,
a call is made to constant-folding to optimize/simplify the address
computation.
The constant-folding, however, makes a mistake in that it sometimes reads
back stale Idxs instead of NewIdxs, that it re-computed in previous iteration.
This leads to incorrect addresses coming out of constant-folding to GEP.
A test case is included. The error is only triggered when indexes have particular
patterns that the stale/new index updates interplay matters.
Reviewers: Daniel Berlin
Differential Revision: https://reviews.llvm.org/D30642
llvm-svn: 297317
While working on improvements to region info analysis, this test case caused an
incorrect region bb2 => bb3 to be detected.
Reviewers: grosser
Contributed-by: Nandini Singhal <cs15mtech01004@iith.ac.in>
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30652
llvm-svn: 297014
While working on improvements to the region info analysis, this test case caused
an incorrect region 1 => 2 to be detected. It is incorrect because entry has an
outgoing edge to 3. This is interesting because 1 dom 2 and 2 pdom 1, which
should have been enough to prevent incoming forward edges into the region and
outgoing forward edges from the region.
Reviewers: grosser
Subscribers: llvm-commits
Contributed-by: Nandini Singhal <cs15mtech01004@iith.ac.in>
Differential Revision: https://reviews.llvm.org/D30603
llvm-svn: 296988
and also "clang-format GenericDomTreeConstruction.h, since the current
formatting makes it look like their is a bug in the loop indentation, and there
is not"
This reverts commit r296535.
There are still some open design questions which I would like to discuss. I
revert this for Daniel (who gave the OK), as he is on vacation.
llvm-svn: 296812
Summary:
Currently, our post-dom tree tries to ignore and remove the effects of
infinite loops. It fails miserably at this, because it tries to do it
ahead of time, and thus can only detect self-loops, and any other type
of infinite loop, it pretends doesn't exist at all.
This can, in a bunch of cases, lead to wrong answers and a completely
empty post-dom tree.
Wrong answer:
```
declare void foo()
define internal void @f() {
entry:
br i1 undef, label %bb35, label %bb3.i
bb3.i:
call void @foo()
br label %bb3.i
bb35.loopexit3:
br label %bb35
bb35:
ret void
}
```
We get:
```
Inorder PostDominator Tree:
[1] <<exit node>> {0,7}
[2] %bb35 {1,6}
[3] %bb35.loopexit3 {2,3}
[3] %entry {4,5}
```
This is a trivial modification of the testcase for PR 6047
Note that we pretend bb3.i doesn't exist.
We also pretend that bb35 post-dominates entry.
While it's true that it does not exit in a theoretical sense, it's not
really helpful to try to ignore the effect and pretend that bb35
post-dominates entry. Worse, we pretend the infinite loop does
nothing (it's usually considered a side-effect), and doesn't even
exist, even when it calls a function. Sadly, this makes it impossible
to use when you are trying to move code safely. All compilers also
create virtual or real single exit nodes (including us), and connect
infinite loops there (which this patch does). In fact, others have
worked around our behavior here, to the point of building their own
post-dom trees:
https://zneak.github.io/fcd/2016/02/17/structuring.html and pointing
out the region infrastructure is near-useless for them with postdom in
this state :(
Completely empty post-dom tree:
```
define void @spam() #0 {
bb:
br label %bb1
bb1: ; preds = %bb1, %bb
br label %bb1
bb2: ; No predecessors!
ret void
}
```
Printing analysis 'Post-Dominator Tree Construction' for function 'foo':
=============================--------------------------------
Inorder PostDominator Tree:
[1] <<exit node>> {0,1}
:(
(note that even if you ignore the effects of infinite loops, bb2
should be present as an exit node that post-dominates nothing).
This patch changes post-dom to properly handle infinite loops and does
root finding during calculation to prevent empty tress in such cases.
We match gcc's (and the canonical theoretical) behavior for infinite
loops (find the backedge, connect it to the exit block).
Testcases coming as soon as i finish running this on a ton of random graphs :)
Reviewers: chandlerc, davide
Subscribers: bryant, llvm-commits
Differential Revision: https://reviews.llvm.org/D29705
llvm-svn: 296535
Summary:
Motivation: fix PR31181 without regression (the actual fix is still in
progress). However, the actual content of PR31181 is not relevant
here.
This change makes poison propagation more aggressive in the following
cases:
1. poision * Val == poison, for any Val. In particular, this changes
existing intentional and documented behavior in these two cases:
a. Val is 0
b. Val is 2^k * N
2. poison << Val == poison, for any Val
3. getelementptr is poison if any input is poison
I think all of these are justified (and are axiomatically true in the
new poison / undef model):
1a: we need poison * 0 to be poison to allow transforms like these:
A * (B + C) ==> A * B + A * C
If poison * 0 were 0 then the above transform could not be allowed
since e.g. we could have A = poison, B = 1, C = -1, making the LHS
poison * (1 + -1) = poison * 0 = 0
and the RHS
poison * 1 + poison * -1 = poison + poison = poison
1b: we need e.g. poison * 4 to be poison since we want to allow
A * 4 ==> A + A + A + A
If poison * 4 were a value with all of their bits poison except the
last four; then we'd not be able to do this transform since then if A
were poison the LHS would only be "partially" poison while the RHS
would be "full" poison.
2: Same reasoning as (1b), we'd like have the following kinds
transforms be legal:
A << 1 ==> A + A
Reviewers: majnemer, efriedma
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D30185
llvm-svn: 295809
Newer ppc supports unaligned memory access, it reduces the cost of unaligned memory access significantly. This patch handles this case in PPCTTIImpl::getMemoryOpCost.
This patch fixes pr31492.
Differential Revision: https://reviews.llvm.org/D28630
This is resubmit of r292680, which was reverted by r293092. The internal application failures were actually caused by a source code bug.
llvm-svn: 295506
proven larger than the loop-count
This fixes PR31098: Try to resolve statically data-dependences whose
compile-time-unknown distance can be proven larger than the loop-count,
instead of resorting to runtime dependence checking (which are not always
possible).
For vectorization it is sufficient to prove that the dependence distance
is >= VF; But in some cases we can prune unknown dependence distances early,
and even before selecting the VF, and without a runtime test, by comparing
the distance against the loop iteration count. Since the vectorized code
will be executed only if LoopCount >= VF, proving distance >= LoopCount
also guarantees that distance >= VF. This check is also equivalent to the
Strong SIV Test.
Reviewers: mkuper, anemet, sanjoy
Differential Revision: https://reviews.llvm.org/D28044
llvm-svn: 294892
Make SolveLinEquationWithOverflow take the start as a SCEV, so we can
solve more cases. With that implemented, get rid of the special case
for powers of two.
The additional functionality probably isn't particularly useful,
but it might help a little for certain cases involving pointer
arithmetic.
Differential Revision: https://reviews.llvm.org/D28884
llvm-svn: 293576
This is fixing pr31761: BasicAA is deducing NoAlias
on the result of the GEP if the base pointer is itself NoAlias.
This is possible only if the NoAlias on the base pointer is
deduced with a non-sized query: this should guarantee that
the pointers are belonging to different memory allocation
and that the GEP can't legally jump from one to another.
Differential Revision: https://reviews.llvm.org/D29216
llvm-svn: 293293
Inlining in getAddExpr() can cause abnormal computational time in some cases.
New parameter -scev-addops-inline-threshold is intruduced with default value 500.
Reviewers: sanjoy
Subscribers: mzolotukhin, llvm-commits
Differential Revision: https://reviews.llvm.org/D28812
llvm-svn: 293176
This reverts commit r292680. It is causing significantly worse
performance and test timeouts in our internal builds. I have already
routed reproduction instructions your way.
llvm-svn: 293092
bots ever since d0k fixed the CHECK lines so that it did something at
all.
It isn't actually testing SCEV directly but LSR, so move it into LSR and
the x86-specific tree of tests that already exists there. Target
dependence is common and unavoidable with the current design of LSR.
llvm-svn: 292774
become unavailable.
The AssumptionCache is now immutable but it still needs to respond to
DomTree invalidation if it ended up caching one.
This lets us remove one of the explicit invalidates of LVI but the
other one continues to avoid hitting a latent bug.
llvm-svn: 292769
Newer ppc supports unaligned memory access, it reduces the cost of unaligned memory access significantly. This patch handles this case in PPCTTIImpl::getMemoryOpCost.
This patch fixes pr31492.
Differential Revision: https://reviews.llvm.org/D28630
llvm-svn: 292680
To avoid regressions, make ScalarEvolution::createSCEV a bit more
clever.
Also get rid of some useless code in ScalarEvolution::howFarToZero
which was hiding this bug.
No new testcase because it's impossible to actually expose this bug:
we don't have any in-tree users of getUDivExactExpr besides the two
functions I just mentioned, and they both dodged the problem. I'll
try to add some interesting users in a followup.
Differential Revision: https://reviews.llvm.org/D28587
llvm-svn: 292449
runnig LCSSA over them prior to running the loop pipeline.
This also teaches the loop PM to verify that LCSSA form is preserved
throughout the pipeline's run across the loop nest.
Most of the test updates just leverage this new functionality. One has to be
relaxed with the new PM as IVUsers is less powerful when it sees LCSSA input.
Differential Revision: https://reviews.llvm.org/D28743
llvm-svn: 292241
First, I've moved a test of IVUsers from the LSR tree to a dedicated
IVUsers test directory. I've also simplified its RUN line now that the
new pass manager's loop PM is providing analyses on their own.
No functionality changed, but it makes subsequent changes cleaner.
llvm-svn: 292060
mark it as never invalidated in the new PM.
The old PM already required this to work, and after a discussion with
Hal this seems to really be the only sensible answer. The cache
gracefully degrades as the IR is mutated, and most things which do this
should already be incrementally updating the cache.
This gets rid of a bunch of logic preserving and testing the
invalidation of this analysis.
llvm-svn: 292039
Refines max backedge-taken count if a loop like
"for (int i = 0; i != n; ++i) { /* body */ }" is rotated.
Differential Revision: https://reviews.llvm.org/D28536
llvm-svn: 291704
This is both easier to understand, and produces a tighter bound in certain
cases.
Differential Revision: https://reviews.llvm.org/D28393
llvm-svn: 291701
updated instructions:
pmulld, pmullw, pmulhw, mulsd, mulps, mulpd, divss, divps, divsd, divpd, addpd and subpd.
special optimization case which replaces pmulld with pmullw\pmulhw\pshuf seq.
In case if the real operands bitwidth <= 16.
Differential Revision: https://reviews.llvm.org/D28104
llvm-svn: 291657
The original code considered only v2i64 as slow for this feature. This patch
consider all 128-bit long vector types as slow candidates.
In internal tests, extending this feature to all 128-bit vector types
resulted in an overall improvement of 1% on Exynos M1.
Differential revision: https://reviews.llvm.org/D27998
llvm-svn: 291616
invalid.
This fixes use-after-free bugs that will arise with any interesting use
of SCEV.
I've added a dedicated test that works diligently to trigger these kinds
of bugs in the new pass manager and also checks for them explicitly as
well as triggering ASan failures when things go squirly.
llvm-svn: 291426
The 'fast' costs should only work for shifts by uniform constants (uniform non-constant are lowered using the slow default implementation).
Logical shifts were not taking into account that we must mask the psrlw result, so the costs needed to be doubled.
Added missing AVX2/AVX512BW costs as well.
llvm-svn: 291391
This test case has been reduced from test/Analysis/RegionInfo/mix_1.ll and
provides us with a minimal example of a test case which caused problems while
working on an improved version of the RegionInfo analysis. We upstream this
test case, as it certainly can be helpful in future debugging and optimization
tests.
Test case reduced by Pratik Bhatu <cs12b1010@iith.ac.in>
llvm-svn: 290974
X86 target does not provide any target specific cost calculation for interleave patterns.It uses the common target-independent calculation, which gives very high numbers. As a result, the scalar version is chosen in many cases. The situation on AVX-512 is even worse, since we have 3-src shuffles that significantly reduce the cost.
In this patch I calculate the cost on AVX-512. It will allow to compare interleave pattern with gather/scatter and choose a better solution (PR31426).
* Shiffle-broadcast cost will be changed in Simon's upcoming patch.
Differential Revision: https://reviews.llvm.org/D28118
llvm-svn: 290810
This fixes the issue exposed in PR31393, where we weren't trying
sufficiently hard to diagnose bad TBAA metadata.
This does reduce the variety in the error messages we print out, but I
think the tradeoff of verifying more, simply and quickly overrules the
need for more helpful error messags here.
llvm-svn: 290713
BasicAA in r290603.
I've kept the basic testing in the new PM test file as that also covers
the AAManager invalidation logic. If/when there is a good place for
broader AA testing it could move there.
This test is somewhat unsatisfying as I can't get it to fail even with
ASan outside of explicit checks of the invalidation. Apparently we don't
yet have any test coverage of the BasicAA code paths using either the
domtree or loopinfo -- I made both of them always be null and check-llvm
passed.
llvm-svn: 290612
For vector GEPs, CastGEPIndices can end up in an infinite recursion, because
we compare the vector type to the scalar pointer type, find them different,
and then try to cast a type to itself.
Differential Revision: https://reviews.llvm.org/D28009
llvm-svn: 290260
This patch checks that the SlowMisaligned128Store subtarget feature is set
when penalizing such stores in getMemoryOpCost.
Differential Revision: https://reviews.llvm.org/D27677
llvm-svn: 289845
After r289755, the AssumptionCache is no longer needed. Variables affected by
assumptions are now found by using the new operand-bundle-based scheme. This
new scheme is more computationally efficient, and also we need much less
code...
llvm-svn: 289756
There was an efficiency problem with how we processed @llvm.assume in
ValueTracking (and other places). The AssumptionCache tracked all of the
assumptions in a given function. In order to find assumptions relevant to
computing known bits, etc. we searched every assumption in the function. For
ValueTracking, that means that we did O(#assumes * #values) work in InstCombine
and other passes (with a constant factor that can be quite large because we'd
repeat this search at every level of recursion of the analysis).
Several of us discussed this situation at the last developers' meeting, and
this implements the discussed solution: Make the values that an assume might
affect operands of the assume itself. To avoid exposing this detail to
frontends and passes that need not worry about it, I've used the new
operand-bundle feature to add these extra call "operands" in a way that does
not affect the intrinsic's signature. I think this solution is relatively
clean. InstCombine adds these extra operands based on what ValueTracking, LVI,
etc. will need and then those passes need only search the users of the values
under consideration. This should fix the computational-complexity problem.
At this point, no passes depend on the AssumptionCache, and so I'll remove
that as a follow-up change.
Differential Revision: https://reviews.llvm.org/D27259
llvm-svn: 289755
Summary:
This change adds some verification in the IR verifier around struct path
TBAA metadata.
Other than some basic sanity checks (e.g. we get constant integers where
we expect constant integers), this checks:
- That by the time an struct access tuple `(base-type, offset)` is
"reduced" to a scalar base type, the offset is `0`. For instance, in
C++ you can't start from, say `("struct-a", 16)`, and end up with
`("int", 4)` -- by the time the base type is `"int"`, the offset
better be zero. In particular, a variant of this invariant is needed
for `llvm::getMostGenericTBAA` to be correct.
- That there are no cycles in a struct path.
- That struct type nodes have their offsets listed in an ascending
order.
- That when generating the struct access path, you eventually reach the
access type listed in the tbaa tag node.
Reviewers: dexonsmith, chandlerc, reames, mehdi_amini, manmanren
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D26438
llvm-svn: 289402
ConstantFolding tried to cast one of the scalar indices to a vector
type. Instead, use the vector type only for the first index (which
is the only one allowed to be a vector) and use its scalar type
otherwise.
Fixes PR31250.
Reviewers: majnemer
Differential Revision: https://reviews.llvm.org/D27389
llvm-svn: 289073
VSX has instructions lxsiwax/lxsdx that can load 32/64 bit value into VSX register cheaply. That patch makes it known to memory cost model, so the vectorization of the test case in pr30990 is beneficial.
Differential Revision: https://reviews.llvm.org/D26713
llvm-svn: 288560
Currently when cost of scalar operations is evaluated the vector type is
used for scalar operations. Patch fixes this issue and fixes evaluation
of the vector operations cost.
Several test showed that vector cost model is too optimistic. It
allowed vectorization of 8 or less add/fadd operations, though scalar
code is faster. Actually, only for 16 or more operations vector code
provides better performance.
Differential Revision: https://reviews.llvm.org/D26277
llvm-svn: 288398
Matt Arsenault