X86 at least is able to use movmsk or kmov to move the mask to the scalar
domain. Then we can just use test instructions to test individual bits.
This is more efficient than extracting each mask element
individually.
I special cased v1i1 to use the previous behavior. This avoids
poor type legalization of bitcast of v1i1 to i1.
I've skipped expandload/compressstore as I think we need to
handle constant masks for those better first.
Many tests end up with duplicate test instructions due to tail
duplication in the branch folding pass. But the same thing
happens when constructing similar code in C. So its not unique
to the scalarization.
Not sure if this lowering code will also be good for other targets,
but we're only testing X86 today.
Differential Revision: https://reviews.llvm.org/D65319
llvm-svn: 367489
This is a prepatory patch for future work on support exit value rewriting in loops with a mixture of computable and non-computable exit counts. The intention is to be "mostly NFC" - i.e. not enable any interesting new transforms - but in practice, there are some small output changes.
The test differences are caused by cases wherewhere getSCEVAtScope can simplify a single entry phi without needing any knowledge of the loop.
llvm-svn: 367485
Reverse the canonicalization of fneg relative to fmul/fdiv. That makes it
easier to implement the transforms (and possibly other fneg transforms) in
1 place because we can always start the pattern match from fneg (either the
legacy binop or the new unop).
There's a secondary practical benefit seen in PR21914 and PR42681:
https://bugs.llvm.org/show_bug.cgi?id=21914https://bugs.llvm.org/show_bug.cgi?id=42681
...hoisting fneg rather than sinking seems to play nicer with LICM in IR
(although this change may expose analysis holes in the other direction).
1. The instcombine test changes show the expected neutral IR diffs from
reversing the order.
2. The reassociation tests show that we were missing an optimization
opportunity to fold away fneg-of-fneg. My reading of IEEE-754 says
that all of these transforms are allowed (regardless of binop/unop
fneg version) because:
"For all other operations [besides copy/abs/negate/copysign], this
standard does not specify the sign bit of a NaN result."
In all of these transforms, we always have some other binop
(fadd/fsub/fmul/fdiv), so we are free to flip the sign bit of a
potential intermediate NaN operand.
(If that interpretation is wrong, then we must already have a bug in
the existing transforms?)
3. The clang tests shouldn't exist as-is, but that's effectively a
revert of rL367149 (the test broke with an extension of the
pre-existing fneg canonicalization in rL367146).
Differential Revision: https://reviews.llvm.org/D65399
llvm-svn: 367447
When vectorizer strips pointers it can eventually end up with
pointers of two different sizes, then SCEV will crash.
Differential Revision: https://reviews.llvm.org/D65480
llvm-svn: 367443
Currently InstCombiner::foldXorOfICmps() bailouts if the
ICMP it wants to invert has extra uses. As it can be seen
in the tests in previous commit, this is super unfortunate,
this is the single pattern that is left non-canonicalized.
We could analyze if we can also invert all the uses if said ICMP
at the same time, thus not bailing out there.
I'm not seeing any nicer alternative.
llvm-svn: 367439
We have some code marks instructions with struct operands as overdefined,
but if the instruction is a call to a function with tracked arguments,
this breaks the assumption that the lattice values of all call sites
are not overdefined and will be replaced by a constant.
This also re-adds the assertion from D65222, with additionally skipping
non-callsite uses. This patch should address the cases reported in which
the assertion fired.
Fixes PR42738.
Reviewers: efriedma, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D65439
llvm-svn: 367430
Summary:
While `-div-rem-pairs` pass can decompose rem in div+rem pair when div-rem pair
is unsupported by target, nothing performs the opposite fold.
We can't do that in InstCombine or DAGCombine since neither of those has access to TTI.
So it makes most sense to teach `-div-rem-pairs` about it.
If we matched rem in expanded form, we know we will be able to place div-rem pair
next to each other so we won't regress the situation.
Also, we shouldn't decompose rem if we matched already-decomposed form.
This is surprisingly straight-forward otherwise.
The original patch was committed in rL367288 but was reverted in rL367289
because it exposed pre-existing RAUW issues in internal data structures
of the pass; those now have been addressed in a previous patch.
https://bugs.llvm.org/show_bug.cgi?id=42673
Reviewers: spatel, RKSimon, efriedma, ZaMaZaN4iK, bogner
Reviewed By: bogner
Subscribers: bogner, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65298
llvm-svn: 367419
Summary:
`DivRemPairs` internally creates two maps:
* {sign, divident, divisor} -> div instruction
* {sign, divident, divisor} -> rem instruction
Then it iterates over rem map, and looks if there is an entry
in div map with the same key. Then depending on some internal logic
it may RAUW rem instruction with something else.
But if that rem instruction is an input to other div/rem,
then it was used as a key in these maps, so the old value (used in key)
is now dandling, because RAUW didn't update those maps.
And we can't even RAUW map keys in general, there's `ValueMap`,
but we don't have a single `Value` as key...
The bug was discovered via D65298, and the test there exists.
Now, i'm not sure how to expose this issue in trunk.
The bug is clearly there if i change the map keys to be `AssertingVH`/`PoisoningVH`,
but i guess this didn't miscompiled anything thus far?
I really don't think this is benin without that patch.
The fix is actually rather straight-forward - instead of trying to somehow
shoe-horn `ValueMap` here (doesn't fit, key isn't just `Value`), or writing a new
`ValueMap` with key being a struct of `Value`s, we can just have an intermediate
data structure - a vector, each entry containing matching `Div, Rem` pair,
and pre-filling it before doing any modifications.
This way we won't need to query map after doing RAUW, so no bug is possible.
Reviewers: spatel, bogner, RKSimon, craig.topper
Reviewed By: spatel
Subscribers: hiraditya, hans, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65451
llvm-svn: 367417
LoopInfo can be easily preserved by passing it to the functions that
modify the CFG (SplitCriticalEdge and MergeBlockIntoPredecessor.
SplitCriticalEdge also preserves LoopSimplify and LCSSA form when when passing in
LoopInfo. The test case shows that we preserve LoopSimplify and
LoopInfo. Adding addPreservedID(LCSSAID) did not preserve LCSSA for some
reason.
Also I am not sure if it is possible to preserve those in the new pass
manager, as they aren't analysis passes.
Reviewers: reames, hfinkel, davide, jdoerfert
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D65137
llvm-svn: 367332
Summary:
This patch extends the use of the OptimizationRemarkEmitter to provide
information about loops that are not fused, and loops that are not eligible for
fusion. In particular, it uses the OptimizationRemarkAnalysis to identify loops
that are not eligible for fusion and the OptimizationRemarkMissed to identify
loops that cannot be fused.
It also reuses the statistics to provide the messages used in the
OptimizationRemarks. This provides common message strings between the
optimization remarks and the statistics.
I would like feedback on this approach, in general. If people are OK with this,
I will flesh out additional remarks in subsequent commits.
Subscribers: hiraditya, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63844
llvm-svn: 367327
Summary:
I have stumbled into this by accident while preparing to extend backend `x s% C ==/!= 0` handling.
While we did happen to handle this fold in most of the cases,
the folding is indirect - we fold `x u% y` to `x & (y-1)` (iff `y` is power-of-two),
or first turn `x s% -y` to `x u% y`; that does handle most of the cases.
But we can't turn `x s% INT_MIN` to `x u% -INT_MIN`,
and thus we end up being stuck with `(x s% INT_MIN) == 0`.
There is no such restriction for the more general fold:
https://rise4fun.com/Alive/IIeS
To be noted, the fold does not enforce that `y` is a constant,
so it may indeed increase instruction count.
This is consistent with what `x u% y`->`x & (y-1)` already does.
I think it makes sense, it's at most one (simple) extra instruction,
while `rem`ainder is really much more un-simple (and likely **very** costly).
Reviewers: spatel, RKSimon, nikic, xbolva00, craig.topper
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65046
llvm-svn: 367322
The code is now in a good enough state to pass the bunch of tests that
I have run (after fixing the bugs), so let's enable it by default.
Differential Revision: https://reviews.llvm.org/D65277
llvm-svn: 367297
test-suite/MultiSource/Benchmarks/DOE-ProxyApps-C/miniGMG broke:
Only PHI nodes may reference their own value!
%sub33 = srem i32 %sub33, %ranks_in_i
This reverts commit r367288.
llvm-svn: 367289
Summary:
While `-div-rem-pairs` pass can decompose rem in div+rem pair when div-rem pair
is unsupported by target, nothing performs the opposite fold.
We can't do that in InstCombine or DAGCombine since neither of those has access to TTI.
So it makes most sense to teach `-div-rem-pairs` about it.
If we matched rem in expanded form, we know we will be able to place div-rem pair
next to each other so we won't regress the situation.
Also, we shouldn't decompose rem if we matched already-decomposed form.
This is surprisingly straight-forward otherwise.
https://bugs.llvm.org/show_bug.cgi?id=42673
Reviewers: spatel, RKSimon, efriedma, ZaMaZaN4iK, bogner
Reviewed By: bogner
Subscribers: bogner, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65298
llvm-svn: 367288
Globals that are associated with globals with type metadata need to appear
in the merged module because they will reference the global's section directly.
Differential Revision: https://reviews.llvm.org/D65312
llvm-svn: 367242
The backend already does this via isNegatibleForFree(),
but we may want to alter the fneg IR canonicalizations
that currently exist, so we need to try harder to fold
fneg in IR to avoid regressions.
llvm-svn: 367227
Summary: As clarified in D53184, volatile load and store do not trap. Therefore, we should remove volatile checks for instructions in `isGuaranteedToTransferExecutionToSuccessor`.
Reviewers: jdoerfert, efriedma, nikic
Reviewed By: nikic
Subscribers: hiraditya, jfb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65375
llvm-svn: 367226
The backend already does this via isNegatibleForFree(),
but we may want to alter the fneg IR canonicalizations
that currently exist, so we need to try harder to fold
fneg in IR to avoid regressions.
llvm-svn: 367194
Summary:
In D62801, new function attribute `willreturn` was introduced. In short, a function with `willreturn` is guaranteed to come back to the call site(more precise definition is in LangRef).
In this patch, willreturn is annotated for LLVM intrinsics.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: jvesely, nhaehnle, sstefan1, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64904
llvm-svn: 367184
The test case from:
https://bugs.llvm.org/show_bug.cgi?id=42771
...shows a ~30x slowdown caused by the awkward loop iteration (rL207302) that is
seemingly done just to avoid invalidating the instruction iterator. We can instead
delay instruction deletion until we reach the end of the block (or we could delay
until we reach the end of all blocks).
There's a test diff here for a degenerate case with llvm.assume that is not
meaningful in itself, but serves to verify this change in logic.
This change probably doesn't result in much overall compile-time improvement
because we call '-instsimplify' as a standalone pass only once in the standard
-O2 opt pipeline currently.
Differential Revision: https://reviews.llvm.org/D65336
llvm-svn: 367173
unreachable loop.
updatePredecessorProfileMetadata in jumpthreading tries to find the
first dominating predecessor block for a PHI value by searching upwards
the predecessor block chain.
But jumpthreading may see some temporary IR state which contains
unreachable bb not being cleaned up. If an unreachable loop happens to
be on the predecessor block chain, keeping chasing the predecessor
block will run into an infinite loop.
The patch fixes it.
Differential Revision: https://reviews.llvm.org/D65310
llvm-svn: 367154
(Y * (1.0 - Z)) + (X * Z) -->
Y - (Y * Z) + (X * Z) -->
Y + Z * (X - Y)
This is part of solving:
https://bugs.llvm.org/show_bug.cgi?id=42716
Factoring eliminates an instruction, so that should be a good canonicalization.
The potential conversion to FMA would be handled by the backend based on target
capabilities.
Differential Revision: https://reviews.llvm.org/D65305
llvm-svn: 367101
Currently there are a few pointer comparisons in ValueDFS_Compare, which
can cause non-deterministic ordering when materializing values. There
are 2 cases this patch fixes:
1. Order defs before uses used to compare pointers, which guarantees
defs before uses, but causes non-deterministic ordering between 2
uses or 2 defs, depending on the allocation order. By converting the
pointers to booleans, we can circumvent that problem.
2. comparePHIRelated was comparing the basic block pointers of edges,
which also results in a non-deterministic order and is also not
really meaningful for ordering. By ordering by their destination DFS
numbers we guarantee a deterministic order.
For the example below, we can end up with 2 different uselist orderings,
when running `opt -mem2reg -ipsccp` hundreds of times. Because the
non-determinism is caused by allocation ordering, we cannot reproduce it
with ipsccp alone.
declare i32 @hoge() local_unnamed_addr #0
define dso_local i32 @ham(i8* %arg, i8* %arg1) #0 {
bb:
%tmp = alloca i32
%tmp2 = alloca i32, align 4
br label %bb19
bb4: ; preds = %bb20
br label %bb6
bb6: ; preds = %bb4
%tmp7 = call i32 @hoge()
store i32 %tmp7, i32* %tmp
%tmp8 = load i32, i32* %tmp
%tmp9 = icmp eq i32 %tmp8, 912730082
%tmp10 = load i32, i32* %tmp
br i1 %tmp9, label %bb11, label %bb16
bb11: ; preds = %bb6
unreachable
bb13: ; preds = %bb20
br label %bb14
bb14: ; preds = %bb13
%tmp15 = load i32, i32* %tmp
br label %bb16
bb16: ; preds = %bb14, %bb6
%tmp17 = phi i32 [ %tmp10, %bb6 ], [ 0, %bb14 ]
br label %bb19
bb18: ; preds = %bb20
unreachable
bb19: ; preds = %bb16, %bb
br label %bb20
bb20: ; preds = %bb19
indirectbr i8* null, [label %bb4, label %bb13, label %bb18]
}
Reviewers: davide, efriedma
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D64866
llvm-svn: 367049
As discussed in https://bugs.llvm.org/show_bug.cgi?id=42673
there is a TTI hook hasDivRemOp() that matters here.
While -div-rem-pairs will decompose 'rem' if that hook returns false,
nothing does the opposite transform.
We can't to this in InstCombine, because it does not currently
access TTI, and i'm not sure we should change that.
We can't really do that in DAGCombine since it also currently does not
access TTI.
Therefore only DivRemPairs is left.
https://bugs.llvm.org/show_bug.cgi?id=42673
llvm-svn: 367046
We'd like to determine the idom of exit block after peeling one iteration.
Let Exit is exit block.
Let ExitingSet - is a set of predecessors of Exit block. They are exiting blocks.
Let Latch' and ExitingSet' are copies after a peeling.
We'd like to find an idom'(Exit) - idom of Exit after peeling.
It is an evident that idom'(Exit) will be the nearest common dominator of ExitingSet and ExitingSet'.
idom(Exit) is a nearest common dominator of ExitingSet.
idom(Exit)' is a nearest common dominator of ExitingSet'.
Taking into account that we have a single Latch, Latch' will dominate Header and idom(Exit).
So the idom'(Exit) is nearest common dominator of idom(Exit)' and Latch'.
All these basic blocks are in the same loop, so what we find is
(nearest common dominator of idom(Exit) and Latch)'.
Reviewers: reames, fhahn
Reviewed By: reames
Subscribers: hiraditya, zzheng, llvm-commits
Differential Revision: https://reviews.llvm.org/D65292
llvm-svn: 367044
Later code in TryToSimplifyUncondBranchFromEmptyBlock() assumes that
we have cleaned up unreachable blocks, but that was not happening
with this switch transform.
llvm-svn: 367037
This reverts commit bc4a63fd3c, this is a
speculative revert to fix a number of sanitizer bots (like
sanitizer-x86_64-linux-bootstrap-ubsan) that have started to see stage2
compiler crashes, presumably due to a miscompile.
llvm-svn: 367029
We do not need the SmallPtrSet to avoid adding duplicates to
OpsToRename, because we already keep a ValueInfo mapping. If we see an
op for the first time, Infos will be empty and we can also add it to
OpsToRename.
We process operands by visiting BBs depth-first and then iterate over
all instructions & users, so the order should be deterministic.
Therefore we can skip one round of sorting, which we purely needed for
guaranteeing a deterministic order when iterating over the SmallPtrSet.
Reviewers: efriedma, davide
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D64816
llvm-svn: 367028
trunc (load X) --> load (bitcast X to narrow type)
We have this transform in DAGCombiner::ReduceLoadWidth(), but the truncated
load pattern can interfere with other instcombine transforms, so I'd like to
allow the fold sooner.
Example:
https://bugs.llvm.org/show_bug.cgi?id=16739
...in that report, we have bitcasts bracketing these ops, so those could get
eliminated too.
We've generally ruled out widening of loads early in IR ( LoadCombine -
http://lists.llvm.org/pipermail/llvm-dev/2016-September/105291.html ), but
that reasoning may not apply to narrowing if we can preserve information
such as the dereferenceable range.
Differential Revision: https://reviews.llvm.org/D64432
llvm-svn: 367011
We can treat icmp eq X, MIN_UINT as icmp ule X, MIN_UINT and allow
it to merge with icmp ugt X, C. Similar for the other constants.
We can do simliar for icmp ne X, (U)INT_MIN/MAX in foldAndOfICmps. And we already handled UINT_MIN there.
Fixes PR42691.
Differential Revision: https://reviews.llvm.org/D65017
llvm-svn: 366945
This is a follow up to D64971. While we need to insert the deref after
the offset, it needs to come before the remaining elements in the
original expression since the deref needs to happen before the LLVM
fragment if present.
Differential Revision: https://reviews.llvm.org/D65172
llvm-svn: 366865
The original code failed to account for the fact that one exit can have a pointer exit count without all of them having pointer exit counts. This could cause two separate bugs:
1) We might exit the loop early, and leave optimizations undone. This is what triggered the assertion failure in the reported test case.
2) We might optimize one exit, then exit without indicating a change. This could result in an analysis invalidaton bug if no other transform is done by the rest of indvars.
Note that the pointer exit counts are a really fragile concept. They show up only when we have a pointer IV w/o a datalayout to provide their size. It's really questionable to me whether the complexity implied is worth it.
llvm-svn: 366829
Craig Topper