While operand bundles carry unpredictable semantics, we know some of
them and can therefore "ignore" them. In this case we allow to look at
the declaration of `llvm.assume` when asked for the attributes at a call
site. The assume operand bundles we have do not invalidate the
declaration attributes.
We cannot test this in isolation because the llvm.assume attributes are
determined by the parser. However, a follow up patch will provide test
coverage.
In `MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4.cpp` we initialized
attributes until stack frame ~35k caused space to run out. The initial
size 1024 is pretty much random.
The get{Return,Unwind,Unreachable}Block functions in
UnifyFunctionExitNodes have not been used for many years,
so just remove them.
Reviewed By: bjope
Differential Revision: https://reviews.llvm.org/D87078
If we know that the abs operand is known negative, we can replace
it with a neg.
To avoid computing known bits twice, I've removed the fold for the
non-negative case from InstSimplify. Both the non-negative and the
negative case are handled by InstCombine now, with one known bits call.
Differential Revision: https://reviews.llvm.org/D87196
D66230 attempted to fix a problem where when there are allocas used before CoroBegin.
It keeps allocas and their uses stay in put if there are no escapse/changes to the data before CoroBegin.
Unfortunately that's incorrect.
Consider this code:
%var = alloca i32
%1 = getelementptr .. %var; stays put
%f = call i8* @llvm.coro.begin
store ... %1
After this fix, %1 will now stay put, however if a store happens after coro.begin and hence modifies the content, this change will not be reflected in the coroutine frame (and will eventually be DCEed).
To generalize the problem, if any alias ptr is created before coro.begin for an Alloca and that alias ptr is latter written into after coro.begin, it will lead to incorrect behavior.
There are also a few other minor issues, such as incorrect dominate condition check in the ptr visitor, unhandled memory intrinsics and etc.
Ths patch attempts to fix some of these issue, and make it more robust to deal with aliases.
While visiting through the alloca pointer, we also keep track of all aliases created that will be used after CoroBegin. We track the offset of each alias, and then reacreate these aliases after CoroBegin using these offset.
It's worth noting that this is not perfect and there will still be cases we cannot handle. I think it's impractical to handle all cases given the current design.
This patch makes it more robust and should be a pure win.
In the meantime, we need to think about what how to completely elimiante these issues, likely through the route as @rjmccall mentioned in D66230.
Differential Revision: https://reviews.llvm.org/D86859
In GenerateConstantOffsetsImpl, we may generate non canonical Formula
if BaseRegs of that Formula is updated and includes a recurrent expr reg
related with current loop while its ScaledReg is not.
Patched by: mdchen
Reviewed By: qcolombet
Differential Revision: https://reviews.llvm.org/D86939
The CloneFunctionInto has implicit requirements with regards to the
linkage and visibility of the function. We now update these after we did
the CloneFunctionInto on the copy with the same linkage and visibility
as the original.
Deleting or replacing anything is certainly a modification. This caused
a later assertion in IPSCCP when compiling 400.perlbench with the new PM.
I'm not sure how to test this.
This was reverted in 503deec218
because it caused gigantic increase (3x) in branch mispredictions
in certain benchmarks on certain CPU's,
see https://reviews.llvm.org/D84108#2227365.
It has since been investigated and here are the results:
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20200907/827578.html
> It's an amazingly severe regression, but it's also all due to branch
> mispredicts (about 3x without this). The code layout looks ok so there's
> probably something else to deal with. I'm not sure there's anything we can
> reasonably do so we'll just have to take the hit for now and wait for
> another code reorganization to make the branch predictor a bit more happy :)
>
> Thanks for giving us some time to investigate and feel free to recommit
> whenever you'd like.
>
> -eric
So let's just reland this.
Original commit message:
I've been looking at missed vectorizations in one codebase.
One particular thing that stands out is that some of the loops
reach vectorizer in a rather mangled form, with weird PHI's,
and some of the loops aren't even in a rotated form.
After taking a more detailed look, that happened because
the loop's headers were too big by then. It is evident that
SimplifyCFG's common code hoisting transform is at fault there,
because the pattern it handles is precisely the unrotated
loop basic block structure.
Surprizingly, `SimplifyCFGOpt::HoistThenElseCodeToIf()` is enabled
by default, and is always run, unlike it's friend, common code sinking
transform, `SinkCommonCodeFromPredecessors()`, which is not enabled
by default and is only run once very late in the pipeline.
I'm proposing to harmonize this, and disable common code hoisting
until //late// in pipeline. Definition of //late// may vary,
here currently i've picked the same one as for code sinking,
but i suppose we could enable it as soon as right after
loop rotation happens.
Experimentation shows that this does indeed unsurprizingly help,
more loops got rotated, although other issues remain elsewhere.
Now, this undoubtedly seriously shakes phase ordering.
This will undoubtedly be a mixed bag in terms of both compile- and
run- time performance, codesize. Since we no longer aggressively
hoist+deduplicate common code, we don't pay the price of said hoisting
(which wasn't big). That may allow more loops to be rotated,
so we pay that price. That, in turn, that may enable all the transforms
that require canonical (rotated) loop form, including but not limited to
vectorization, so we pay that too. And in general, no deduplication means
more [duplicate] instructions going through the optimizations. But there's still
late hoisting, some of them will be caught late.
As per benchmarks i've run {F12360204}, this is mostly within the noise,
there are some small improvements, some small regressions.
One big regression i saw i fixed in rG8d487668d09fb0e4e54f36207f07c1480ffabbfd, but i'm sure
this will expose many more pre-existing missed optimizations, as usual :S
llvm-compile-time-tracker.com thoughts on this:
http://llvm-compile-time-tracker.com/compare.php?from=e40315d2b4ed1e38962a8f33ff151693ed4ada63&to=c8289c0ecbf235da9fb0e3bc052e3c0d6bff5cf9&stat=instructions
* this does regress compile-time by +0.5% geomean (unsurprizingly)
* size impact varies; for ThinLTO it's actually an improvement
The largest fallout appears to be in GVN's load partial redundancy
elimination, it spends *much* more time in
`MemoryDependenceResults::getNonLocalPointerDependency()`.
Non-local `MemoryDependenceResults` is widely-known to be, uh, costly.
There does not appear to be a proper solution to this issue,
other than silencing the compile-time performance regression
by tuning cut-off thresholds in `MemoryDependenceResults`,
at the cost of potentially regressing run-time performance.
D84609 attempts to move in that direction, but the path is unclear
and is going to take some time.
If we look at stats before/after diffs, some excerpts:
* RawSpeed (the target) {F12360200}
* -14 (-73.68%) loops not rotated due to the header size (yay)
* -272 (-0.67%) `"Number of live out of a loop variables"` - good for vectorizer
* -3937 (-64.19%) common instructions hoisted
* +561 (+0.06%) x86 asm instructions
* -2 basic blocks
* +2418 (+0.11%) IR instructions
* vanilla test-suite + RawSpeed + darktable {F12360201}
* -36396 (-65.29%) common instructions hoisted
* +1676 (+0.02%) x86 asm instructions
* +662 (+0.06%) basic blocks
* +4395 (+0.04%) IR instructions
It is likely to be sub-optimal for when optimizing for code size,
so one might want to change tune pipeline by enabling sinking/hoisting
when optimizing for size.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D84108
This reverts commit 503deec218.
For intrinsics supported by ConstantRange, compute the result range
based on the argument ranges. We do this independently of whether
some or all of the input ranges are full, as we can often still
constrain the result in some way.
Differential Revision: https://reviews.llvm.org/D87183
This was supposed to be an NFC cleanup, but there's
a real logic difference (did not drop 'nsw') visible
in some tests in addition to an efficiency improvement.
This is because in the case where we have 2 GEPs,
the code was *always* swapping the operands and
negating the result. But if we have 2 GEPs, we
should *never* need swapping/negation AFAICT.
This is part of improving flags propagation noticed
with PR47430.
Normal dead code elimination ignores assume intrinsics, so we fail to
delete assumes that are not meaningful (and potentially worse if they
cause conflicts with other assumptions).
The motivating example in https://llvm.org/PR47416 suggests that we
might have problems upstream from here (difference between C and C++),
but this should be a cheap way to make sure we remove more dead code.
Differential Revision: https://reviews.llvm.org/D87149
To enable the cost of constants, the helper function has been
reorganised:
- A struct has been introduced to hold SCEV operand information so
that we know the user of the operand, as well as the operand index.
The Worklist now uses instead instead of a bare SCEV.
- The costing of each SCEV, and collection of its operands, is now
performed in a helper function.
Differential Revision: https://reviews.llvm.org/D86050
Modify FoldBranchToCommonDest to consider the cost of inserting
instructions when attempting to combine predicates to fold blocks.
The threshold can be controlled via a new option:
-simplifycfg-branch-fold-threshold which defaults to '2' to allow
the insertion of a not and another logical operator.
Differential Revision: https://reviews.llvm.org/D86526
Similar to D87168, but for abs. If we have a dominating x >= 0
condition, then we know that abs(x) is x. This fold is in
InstCombine, because we need to create a sub instruction for
the x < 0 case.
Differential Revision: https://reviews.llvm.org/D87184
When a switch case is folded into default's case, that's an IR change that
should be reported, update ConstantFoldTerminator accordingly.
Differential Revision: https://reviews.llvm.org/D87142
When checking call sites, give special handling to indirect call, as the
callee may be unknown and can lead to nullptr dereference later. Assume
conservatively that the ICV always changes in such case.
Reviewed By: sstefan1
Differential Revision: https://reviews.llvm.org/D87104
This patch updates MemCpyOpt to preserve MemorySSA. It uses the
MemoryDef at the insertion point of the builder and inserts the new def
after that def.
In some cases, we just modify a memory instruction. In that case, get
the defining access, then remove the memory access and add a new one.
If the defining access is in a different block, insert a new def at the
beginning of the current block, otherwise after the defining access.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D86651
As far as I am aware, the placement of MergedLoadStoreMotion in the
pipeline is not heavily tuned currently. It seems to not matter much if
we do it after DSE in the LTO pipeline (no binary changes for -O3 -flto
on MultiSource/SPEC2000/SPEC2006). Moving it after DSE however has a
major benefit: MemorySSA is constructed by LICM and is consumed by DSE,
so if MergedLoadStoreMotion happens after DSE, we do not need to
preserve MemorySSA in it.
If there are any concerns with this move, I can also update
MergedLoadStoreMotion to preserve MemorySSA.
This patch together with D86651 (preserve MemSSA in MemCpyOpt) and
D86534 (preserve MemSSA in GVN) are the remaining patches to bring down
compile-time for DSE + MemorySSA to the levels outlined in
http://lists.llvm.org/pipermail/llvm-dev/2020-August/144417.html
Once they land, we should be able to start with flipping the switch on
enabling DSE + MmeorySSA.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D86967
Preserve MemorySSA if it is available before running GVN.
DSE with MemorySSA will run closely after GVN. If GVN and 2 other
passes preserve MemorySSA, DSE can re-use MemorySSA used by LICM
when doing LTO.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D86534
This is a followup to 1ccfb52a61, which made a number of changes
including the apparently innocuous reordering of required passes in
MemCpyOptimizer. This however altered the creation order of BasicAA vs
Phi Values analysis, meaning BasicAA did not pick up PhiValues as a
cached result. Instead if we require MemoryDependence first it will
require PhiValuesAnalysis allowing BasicAA to use it for better results.
I don't claim this is an excellent design, but it fixes a nasty little
regressions where a query later in JumpThreading was getting worse
results.
Differential Revision: https://reviews.llvm.org/D87027
Currently IPSCCP (and others like CVP/GVN) blindly propagate pointer
equalities. In certain cases, that leads to dereferenceable pointers
being replaced, as in the example test case.
I think this is not allowed, as it introduces an access of an
un-dereferenceable pointer. Note that the pointer is inbounds, but one
past the last element, so it is valid, but not dereferenceable.
This patch is mostly to highlight the issue and start a discussion.
Currently it only checks for specifically looking
one-past-the-last-element pointers with array typed bases.
This causes the mis-compile outlined in
https://stackoverflow.com/questions/55754313/is-this-gcc-clang-past-one-pointer-comparison-behavior-conforming-or-non-standar
In the test case, if we replace %p with the GEP for the store, we
subsequently determine that the store and the load cannot alias, because
they are to different underlying objects.
Note that Alive2 seems to think that the replacement is valid:
https://alive2.llvm.org/ce/z/2rorhk
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D85332
First, shuffle cost for scalable type is not known for scalable type;
Second, we cannot reason if the narrowed shuffle mask for scalable type
is a splat or not.
E.g., Bitcast splat vector from type <vscale x 4 x i32> to <vscale x 8 x i16>
will involve narrowing shuffle mask <vscale x 4 x i32> zeroinitializer to
<vscale x 8 x i32> with element sequence of <0, 1, 0, 1, ...>, which cannot be
reasoned if it's a valid splat or not.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D86995
In IPSCCP when a function is optimized to return undef, it should clear the returned attribute for all its input arguments
and its corresponding call sites.
The bug is exposed when the value of an input argument of the function is assigned to a physical register and
because of the argument having a returned attribute, the value of this physical register will continue to be used
as the function return value right after the call instruction returns, even if the value that this register holds may
be clobbered during the function call. This potentially results in incorrect values being used afterwards.
Reviewed By: jdoerfert, fhahn
Differential Revision: https://reviews.llvm.org/D84220
When marking a global variable constant, and simplifying users using
CleanupConstantGlobalUsers(), the pass could incorrectly return false if
there were still some uses left, and no further optimizations was done.
This was caught using the check introduced by D80916.
This fixes PR46749.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D85837
These transforms will now be performed irrespective of the number of uses for the expression "1.0/sqrt(X)":
1.0/sqrt(X) * X => X/sqrt(X)
X * 1.0/sqrt(X) => X/sqrt(X)
We already handle more general cases, and we are intentionally not creating extra (and likely expensive)
fdiv ops in IR. This pattern is the exception to the rule because we always expect the Backend to reduce
X/sqrt(X) to sqrt(X), if it has the necessary (reassoc) fast-math-flags.
Ref: DagCombiner optimizes the X/sqrt(X) to sqrt(X).
Differential Revision: https://reviews.llvm.org/D86726
This is an enhancement to D81766 to allow loading the minimum target
vector type into an IR vector with a different number of elements.
In one of the motivating tests from PR16739, SLP creates <2 x float>
load ops mixed with <4 x float> insert ops, so we want to handle that
pattern in addition to potential oversized vectors created by the
vectorizers.
For now, we are assuming the insert/extract subvector with undef is
free because there is no exact corresponding TTI modeling for that.
Differential Revision: https://reviews.llvm.org/D86160
This patch makes it possible for AAUB to use information from AANoUndef.
This is the next patch of D86983
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D86984
Johannes Doerfert