In rL300494 there was an attempt to deal with excessive compile time on
invocations of getSign/ZeroExtExpr using local caching. This approach only
helps if we request the same SCEV multiple times throughout recursion. But
in the bug PR33431 we see a case where we request different values all the time,
so caching does not help and the size of the cache grows enormously.
In this patch we remove the local cache for this methods and add the recursion
depth limit instead, as we do for arithmetics. This gives us a guarantee that the
invocation sequence is limited and reasonably short.
Differential Revision: https://reviews.llvm.org/D34273
llvm-svn: 306785
Summary:
DFS InOut numbers currently get eagerly computer upon DomTree construction. They are only needed to answer dome dominance queries and they get invalidated by updates and recalculations. Because of that, it is faster in practice to compute them lazily when they are actually needed.
Clang built without this patch takes 6m 45s to boostrap on my machine, and with the patch applied 6m 38s.
Reviewers: sanjoy, dberlin, chandlerc
Reviewed By: dberlin
Subscribers: davide, llvm-commits
Differential Revision: https://reviews.llvm.org/D34296
llvm-svn: 306778
In LLVM IR the following code:
%r = urem <ty> %t, %b
is equivalent to:
%q = udiv <ty> %t, %b
%s = mul <ty> nuw %q, %b
%r = sub <ty> nuw %t, %q ; (t / b) * b + (t % b) = t
As UDiv, Mul and Sub are already supported by SCEV, URem can be
implemented with minimal effort this way.
Note: While SRem and SDiv are also related this way, SCEV does not
provides SDiv yet.
llvm-svn: 306695
The cost of an interleaved access was only implemented for AVX512. For other
X86 targets an overly conservative Base cost was returned, resulting in
avoiding vectorization where it is actually profitable to vectorize.
This patch starts to add costs for AVX2 for most prominent cases of
interleaved accesses (stride 3,4 chars, for now).
Note1: Improvements of up to ~4x were observed in some of EEMBC's rgb
workloads; There is also a known issue of 15-30% degradations on some of these
workloads, associated with an interleaved access followed by type
promotion/widening; the resulting shuffle sequence is currently inefficient and
will be improved by a series of patches that extend the X86InterleavedAccess pass
(such as D34601 and more to follow).
Note 2: The costs in this patch do not reflect port pressure penalties which can
be very dominant in the case of interleaved accesses since most of the shuffle
operations are restricted to a single port. Further tuning, that may incorporate
these considerations, will be done on top of the upcoming improved shuffle
sequences (that is, along with the abovementioned work to extend
X86InterleavedAccess pass).
Differential Revision: https://reviews.llvm.org/D34023
llvm-svn: 306238
Using various methods, BasicAA tries to determine whether two
GetElementPtr memory locations alias when its base pointers are known
to be equal. When none of its heuristics are applicable, it falls back
to PartialAlias to, according to a comment, protect TBAA making a wrong
decision in case of unions and malloc. PartialAlias is not correct,
because a PartialAlias result implies that some, but not all, bytes
overlap which is not necessarily the case here.
AAResults returns the first analysis result that is not MayAlias.
BasicAA is always the first alias analysis. When it returns
PartialAlias, no other analysis is queried to give a more exact result
(which was the intention of returning PartialAlias instead of MayAlias).
For instance, ScopedAA could return a more accurate result.
The PartialAlias hack was introduced in r131781 (and re-applied in
r132632 after some reverts) to fix llvm.org/PR9971 where TBAA returns a
wrong NoAlias result due to a union. A test case for the malloc case
mentioned in the comment was not provided and I don't think it is
affected since it returns an omnipotent char anyway.
Since r303851 (https://reviews.llvm.org/D33328) clang does emit specific
TBAA for unions anymore (but "omnipotent char" instead). Hence, the
PartialAlias workaround is not required anymore.
This patch passes the test-suite and check-llvm/check-clang of a
self-hoisted build on x64.
Reviewed By: hfinkel
Differential Revision: https://reviews.llvm.org/D34318
llvm-svn: 305938
Summary:
After a single predecessor is merged into a basic block, we need to invalidate
the LVI information for the new merged block, when LVI is not provably true for
all of instructions in the new block.
The test cases added show the correct LVI information using the LVI printer
pass.
Reviewers: reames, dberlin, davide, sanjoy
Reviewed by: dberlin, davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D34108
llvm-svn: 305699
This is a fix for PR33292 that shows a case of extremely long compilation
of a single .c file with clang, with most time spent within SCEV.
We have a mechanism of limiting recursion depth for getAddExpr to avoid
long analysis in SCEV. However, there are calls from getAddExpr to getMulExpr
and back that do not propagate the info about depth. As result of this, a chain
getAddExpr -> ... .> getAddExpr -> getMulExpr -> getAddExpr -> ... -> getAddExpr
can be extremely long, with every segment of getAddExpr's being up to max depth long.
This leads either to long compilation or crash by stack overflow. We face this situation while
analyzing big SCEVs in the test of PR33292.
This patch applies the same limit on max expression depth for getAddExpr and getMulExpr.
Differential Revision: https://reviews.llvm.org/D33984
llvm-svn: 305463
The zero heuristic assumes that integers are more likely positive than negative,
but this also has the effect of assuming that strcmp return values are more
likely positive than negative. Given that for nonzero strcmp return values it's
the ordering of arguments that determines the sign of the result there's no
reason to assume that's true.
Fix this by inspecting the LHS of the compare and using TargetLibraryInfo to
decide if it's strcmp-like, and if so only assume that nonzero is more likely
than zero i.e. strings are more often different than the same. This causes a
slight code generation change in the spec2006 benchmark 403.gcc, but with no
noticeable performance impact. The intent of this patch is to allow better
optimisation of dhrystone on Cortex-M cpus, but currently it won't as there are
also some changes that need to be made to if-conversion.
Differential Revision: https://reviews.llvm.org/D33934
llvm-svn: 304970
Summary:
LVIPrinter pass was previously relying on the LVICache. We now directly call the
the LVI functions which solves the value if the LVI information is not already
available in the cache. This has 2 benefits over the printing of LVI cache:
1. higher coverage (i.e. catches errors) in LVI code when cache value is
invalidated.
2. relies on the core functions, and not dependent on the LVI cache (which may
be scrapped at some point).
It would still catch any cache invalidation errors, since we first go through
the cache.
Reviewers: reames, dberlin, sanjoy
Reviewed by: reames
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D32135
llvm-svn: 304819
Thanks to Galina Kistanova for finding the missing break!
When trying to make a test for this, I realized our logic for handling
extractvalue/insertvalue/... is somewhat broken. This makes constructing
a test-case for this missing break nontrivial.
llvm-svn: 304275
The optimistic delinearization implemented in LLVM detects array sizes by
looking for non-linear products between parameters and induction variables.
In OpenCL code, such products often look like:
A[get_global_id(0) * N + get_global_id(1)]
Hence, the IV is hidden in the get_global_id() call and consequently
delinearization would fail as no induction variable is available that helps
us to identify N as array size parameter.
We now use a very simple heuristic to change this. We assume that each parameter
that comes directly from a function call is a hidden induction variable. As
a result, we can delinearize the access above to:
A[get_global_id(0)][get_global_id(1]
llvm-svn: 304073
The patch rL303730 was reverted because test lsr-expand-quadratic.ll failed on
many non-X86 configs with this patch. The reason of this is that the patch
makes a correctless fix that changes optimizer's behavior for this test.
Without the change, LSR was making an overconfident simplification basing on a
wrong SCEV. Apparently it did not need the IV analysis to do this. With the
change, it chose a different way to simplify (that wasn't so confident), and
this way required the IV analysis. Now, following the right execution path,
LSR tries to make a transformation relying on IV Users analysis. This analysis
is target-dependent due to this code:
// LSR is not APInt clean, do not touch integers bigger than 64-bits.
// Also avoid creating IVs of non-native types. For example, we don't want a
// 64-bit IV in 32-bit code just because the loop has one 64-bit cast.
uint64_t Width = SE->getTypeSizeInBits(I->getType());
if (Width > 64 || !DL.isLegalInteger(Width))
return false;
To make a proper transformation in this test case, the type i32 needs to be
legal for the specified data layout. When the test runs on some non-X86
configuration (e.g. pure ARM 64), opt gets confused by the specified target
and does not use it, rejecting the specified data layout as well. Instead,
it uses some default layout that does not treat i32 as a legal type
(currently the layout that is used when it is not specified does not have
legal types at all). As result, the transformation we expect to happen does
not happen for this test.
This re-enabling patch does not have any source code changes compared to the
original patch rL303730. The only difference is that the failing test is
moved to X86 directory and now has requirement of running on x86 only to comply
with the specified target triple and data layout.
Differential Revision: https://reviews.llvm.org/D33543
llvm-svn: 303971
When folding arguments of AddExpr or MulExpr with recurrences, we rely on the fact that
the loop of our base recurrency is the bottom-lost in terms of domination. This assumption
may be broken by an expression which is treated as invariant, and which depends on a complex
Phi for which SCEVUnknown was created. If such Phi is a loop Phi, and this loop is lower than
the chosen AddRecExpr's loop, it is invalid to fold our expression with the recurrence.
Another reason why it might be invalid to fold SCEVUnknown into Phi start value is that unlike
other SCEVs, SCEVUnknown are sometimes position-bound. For example, here:
for (...) { // loop
phi = {A,+,B}
}
X = load ...
Folding phi + X into {A+X,+,B}<loop> actually makes no sense, because X does not exist and cannot
exist while we are iterating in loop (this memory can be even not allocated and not filled by this moment).
It is only valid to make such folding if X is defined before the loop. In this case the recurrence {A+X,+,B}<loop>
may be existant.
This patch prohibits folding of SCEVUnknown (and those who use them) into the start value of an AddRecExpr,
if this instruction is dominated by the loop. Merging the dominating unknown values is still valid. Some tests that
relied on the fact that some SCEVUnknown should be folded into AddRec's are changed so that they no longer
expect such behavior.
llvm-svn: 303730
This is a re-application of a r303497 that was reverted in r303498.
I thought it had broken a bot when it had not (the breakage did not
go away with the revert).
This change makes the split between the "exact" backedge taken count
and the "maximum" backedge taken count a bit more obvious. Both of
these are upper bounds on the number of times the loop header
executes (since SCEV does not account for most kinds of abnormal
control flow), but the latter is guaranteed to be a constant.
There were a few places where the max backedge taken count *was* a
non-constant; I've changed those to compute constants instead.
At this point, I'm not sure if the constant max backedge count can be
computed by calling `getUnsignedRange(Exact).getUnsignedMax()` without
losing precision. If it can, we can simplify even further by making
`getMaxBackedgeTakenCount` a thin wrapper around
`getBackedgeTakenCount` and `getUnsignedRange`.
llvm-svn: 303531
This change makes the split between the "exact" backedge taken count
and the "maximum" backedge taken count a bit more obvious. Both of
these are upper bounds on the number of times the loop header
executes (since SCEV does not account for most kinds of abnormal
control flow), but the latter is guaranteed to be a constant.
There were a few places where the max backedge taken count *was* a
non-constant; I've changed those to compute constants instead.
At this point, I'm not sure if the constant max backedge count can be
computed by calling `getUnsignedRange(Exact).getUnsignedMax()` without
losing precision. If it can, we can simplify even further by making
`getMaxBackedgeTakenCount` a thin wrapper around
`getBackedgeTakenCount` and `getUnsignedRange`.
llvm-svn: 303497
The probability of edge coming to unreachable block should be as low as possible.
The change reduces the probability to minimal value greater than zero.
The bug https://bugs.llvm.org/show_bug.cgi?id=32214 show the example when
the probability of edge coming to unreachable block is greater than for edge
coming to out of the loop and it causes incorrect loop rotation.
Please note that with this change the behavior of unreachable heuristic is a bit different
than others. Specifically, before this change the sum of probabilities
coming to unreachable blocks have the same weight for all branches
(it was just split over all edges of this block coming to unreachable blocks).
With this change it might be slightly different but not to much due to probability of
taken branch to unreachable block is really small.
Reviewers: chandlerc, sanjoy, vsk, congh, junbuml, davidxl, dexonsmith
Reviewed By: chandlerc, dexonsmith
Subscribers: reames, llvm-commits
Differential Revision: https://reviews.llvm.org/D30633
llvm-svn: 303327
The existing sorting order in defined CompareSCEVComplexity sorts AddRecExprs
by loop depth, but does not pay attention to dominance of loops. This can
lead us to the following buggy situation:
for (...) { // loop1
op1 = {A,+,B}
}
for (...) { // loop2
op2 = {A,+,B}
S = add op1, op2
}
In this case there is no guarantee that in operand list of S the op2 comes
before op1 (loop depth is the same, so they will be sorted just
lexicographically), so we can incorrectly treat S as a recurrence of loop1,
which is wrong.
This patch changes the sorting logic so that it places the dominated recs
before the dominating recs. This ensures that when we pick the first recurrency
in the operands order, it will be the bottom-most in terms of domination tree.
The attached test set includes some tests that produce incorrect SCEV
estimations and crashes with oldlogic.
Reviewers: sanjoy, reames, apilipenko, anna
Reviewed By: sanjoy
Subscribers: llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D33121
llvm-svn: 303148
Update a few tests to use llvm.masked.load/store instead of arm neon
vector loads and stores, and move the tests that are actually specific
to those arm intrinsics to their own files. This lets us mark the
tests that use target specific intrinsics as requiring those targets.
llvm-svn: 302972
This is a follow up patch for https://reviews.llvm.org/rL300440
to address a comment.
To make implementation to be consistent with other cases we just
ignore the remainder after distribution of remaining probability between
reachable edges.
If we reduced the probability of some edges coming to unreachable
blocks we should distribute the remaining part across other edges
coming to reachable blocks to satisfy the condition that sum of all
probabilities should be equal to one. If this remaining part is not
divided by number of "reachable" edges then we get this remainder.
This remainder probability should be pretty small. Other cases just ignore
if the sum of probabilities is not equal to one so we do the same.
Reviewers: chandlerc, sanjoy, vsk, junbuml, reames
Reviewed By: reames
Subscribers: reames, llvm-commits
Differential Revision: https://reviews.llvm.org/D32124
llvm-svn: 302883
The AArch64 instruction set has a few "widening" instructions (e.g., uaddl,
saddl, uaddw, etc.) that take one or more doubleword operands and produce
quadword results. The operands are automatically sign- or zero-extended as
appropriate. However, in LLVM IR, these extends are explicit. This patch
updates TTI to consider these widening instructions as single operations whose
cost is attached to the arithmetic instruction. It marks extends that are part
of a widening operation "free" and applies a sub-target specified overhead
(zero by default) to the arithmetic instructions.
Differential Revision: https://reviews.llvm.org/D32706
llvm-svn: 302582
Account for subvector extraction/insertion, helps prevent the vectorizers from selecting 256-bit vectors that will have to be split anyhow on AVX1 targets.
llvm-svn: 302378
Summary:
The existing implementation creates a symbolic SCEV expression every
time we analyze a phi node and then has to remove it, when the analysis
is finished. This is very expensive, and in most of the cases it's also
unnecessary. According to the data I collected, ~60-70% of analyzed phi
nodes (measured on SPEC) have the following form:
PN = phi(Start, OP(Self, Constant))
Handling such cases separately significantly speeds this up.
Reviewers: sanjoy, pete
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D32663
llvm-svn: 302096
Fixes PR31789 - When loop-vectorize tries to use these intrinsics for a
non-default address space pointer we fail with a "Calling a function with a
bad singature!" assertion. This patch solves this by adding the 'vector of
pointers' argument as an overloaded type which will determine the address
space.
Differential revision: https://reviews.llvm.org/D31490
llvm-svn: 302018
In cases where an instruction (a call site, say) is RAUW'ed with some
other value (this is possible via the `returned` attribute, for
instance), we want the slot in UnknownInsts to point to the original
Instruction we wanted to track, not the value it got replaced by.
Fixes PR32587.
This relands r301426.
llvm-svn: 301814
Summary:
programUndefinedIfPoison makes more sense, given what the function
does; and I'm about to add a function with a name similar to
isKnownNotFullPoison (so do the rename to avoid confusion).
Reviewers: broune, majnemer, bjarke.roune
Reviewed By: broune
Subscribers: mcrosier, llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D30444
llvm-svn: 301776
Commits were:
"Use WeakVH instead of WeakTrackingVH in AliasSetTracker's UnkownInsts"
"Add a new WeakVH value handle; NFC"
"Rename WeakVH to WeakTrackingVH; NFC"
The changes assumed pointers are 8 byte aligned on all architectures.
llvm-svn: 301429
Max Kazantsev