I've amended the isLoadInvariantInLoop function to bail out for
scalable vectors for now since the invariant.start intrinsic is only
ever generated by the clang frontend for thread locals or struct
and class constructors, neither of which support sizeless types.
In addition, the intrinsic itself does not currently support the
concept of a scaled size, which makes it impossible to compare
the sizes of different scalable objects, e.g. <vscale x 32 x i8>
and <vscale x 16 x i8>.
Added new tests here:
Transforms/LICM/AArch64/sve-load-hoist.ll
Transforms/LICM/hoisting.ll
Differential Revision: https://reviews.llvm.org/D87227
AliasAnalysis/MemoryLocation does not account for loops. Two
MemoryLocation can be must-overwrite, even if the first one writes
multiple locations in a loop.
This patch prevents removing such stores, by only considering candidates
that are known to be loop invariant, or executed in the same BB.
Currently the invariant check is quite conservative and only considers
Alloca and Alloca-like instructions and arguments as invariant base pointers.
It also considers GEPs with all constant indices and invariant bases as
invariant.
This can be improved in the future, but the current implementation has
only minor impact on the total number of stores eliminated (25903 vs
26047 for the baseline). There are some 2-10% swings for some individual
benchmarks. In roughly half of the cases, the number of stores removed
increases actually, because we skip candidates that are unlikely to be
valid candidates early.
The function LoopIdiomRecognize::isLegalStore looks for stores in loops
that could be transformed into memset or memcpy. However, the algorithm
currently requires that we know how big the store is at runtime, i.e.
that the store size will not overflow an unsigned integer. For scalable
vectors we cannot guarantee this so I have changed the code to bail out
for now. In addition, even if we add a way to query the maximum value of
vscale in future we will still need to update the algorithm to cope with
non-constant strides. The additional cost associated with calculating
the memset and memcpy arguments will need to be taken into account as
well.
This patch also fixes up an implicit TypeSize -> uint64_t cast,
thereby removing a warning. I've added tests here showing a fixed
width vector loop being transformed into memcpy, and a scalable
vector loop remaining unchanged:
Transforms/LoopIdiom/memcpy-vectors.ll
Differential Revision: https://reviews.llvm.org/D87439
This fixes PR47297.
When ProcessBlock() was able to constant fold the terminator's
condition, but not do any more transformations, the function would
return false, which would lead to the JumpThreading pass returning an
incorrect modified status. This patch makes so that ProcessBlock()
returns true in such cases. This will trigger an unnecessary invocation
of ProcessBlock() in such cases, but this should be rare to occur.
This was caught using the check introduced by D80916.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D87392
When deleting stores at the end of a function, we have to do PHI
translation, otherwise we might miss reads in different iterations of a
loop. See multiblock-loop-carried-dependence.ll for details.
This fixes a mis-compile and surprisingly also increases the number of
eliminated stores from 26047 to 26572 for MultiSource/SPEC2000/SPEC2006
on X86 with -O3 -flto. This is most likely because we save budget by not
exploring through MemoryPhis, which are less likely to result in valid
candidates for elimination.
The issue was reported post-commit for fb109c42d9.
NOTE: There is a mailing list discussion on this: http://lists.llvm.org/pipermail/llvm-dev/2019-December/137632.html
Complemantary to the assumption outliner prototype in D71692, this patch
shows how we could simplify the code emitted for an alignemnt
assumption. The generated code is smaller, less fragile, and it makes it
easier to recognize the additional use as a "assumption use".
As mentioned in D71692 and on the mailing list, we could adopt this
scheme, and similar schemes for other patterns, without adopting the
assumption outlining.
In addition to calculate hash consistently by swapping SELECT's
operands, we also need to inverse the select pattern favor to match the
original logic.
[EarlyCSE] Equivalent SELECTs should hash equally
DenseMap<SimpleValue> assumes that, if its isEqual method returns true
for two elements, then its getHashValue method must return the same value
for them. This invariant is broken when one SELECT node is a min/max
operation, and the other can be transformed into an equivalent min/max by
inverting its predicate and swapping its operands. This patch fixes an
assertion failure that would occur intermittently while compiling the
following IR:
define i32 @t(i32 %i) {
%cmp = icmp sle i32 0, %i
%twin1 = select i1 %cmp, i32 %i, i32 0
%cmpinv = icmp sgt i32 0, %i
%twin2 = select i1 %cmpinv, i32 0, i32 %i
%sink = add i32 %twin1, %twin2
ret i32 %sink
}
Differential Revision: https://reviews.llvm.org/D86843
The tests have been updated and I plan to move them from the MSSA
directory up.
Some end-to-end tests needed small adjustments. One difference to the
legacy DSE is that legacy DSE also deletes trivially dead instructions
that are unrelated to memory operations. Because MemorySSA-backed DSE
just walks the MemorySSA, we only visit/check memory instructions. But
removing unrelated dead instructions is not really DSE's job and other
passes will clean up.
One noteworthy change is in llvm/test/Transforms/Coroutines/ArgAddr.ll,
but I think this comes down to legacy DSE not handling instructions that
may throw correctly in that case. To cover this with MemorySSA-backed
DSE, we need an update to llvm.coro.begin to treat it's return value to
belong to the same underlying object as the passed pointer.
There are some minor cases MemorySSA-backed DSE currently misses, e.g. related
to atomic operations, but I think those can be implemented after the switch.
This has been discussed on llvm-dev:
http://lists.llvm.org/pipermail/llvm-dev/2020-August/144417.html
For the MultiSource/SPEC2000/SPEC2006 the number of eliminated stores
goes from ~17500 (legayc DSE) to ~26300 (MemorySSA-backed). More numbers
and details in the thread on llvm-dev.
Impact on CTMark:
```
Legacy Pass Manager
exec instrs size-text
O3 + 0.60% - 0.27%
ReleaseThinLTO + 1.00% - 0.42%
ReleaseLTO-g. + 0.77% - 0.33%
RelThinLTO (link only) + 0.87% - 0.42%
RelLO-g (link only) + 0.78% - 0.33%
```
http://llvm-compile-time-tracker.com/compare.php?from=3f22e96d95c71ded906c67067d75278efb0a2525&to=ae8be4642533ff03803967ee9d7017c0d73b0ee0&stat=instructions
```
New Pass Manager
exec instrs. size-text
O3 + 0.95% - 0.25%
ReleaseThinLTO + 1.34% - 0.41%
ReleaseLTO-g. + 1.71% - 0.35%
RelThinLTO (link only) + 0.96% - 0.41%
RelLO-g (link only) + 2.21% - 0.35%
```
http://195.201.131.214:8000/compare.php?from=3f22e96d95c71ded906c67067d75278efb0a2525&to=ae8be4642533ff03803967ee9d7017c0d73b0ee0&stat=instructions
Reviewed By: asbirlea, xbolva00, nikic
Differential Revision: https://reviews.llvm.org/D87163
DenseMap<SimpleValue> assumes that, if its isEqual method returns true
for two elements, then its getHashValue method must return the same value
for them. This invariant is broken when one SELECT node is a min/max
operation, and the other can be transformed into an equivalent min/max by
inverting its predicate and swapping its operands. This patch fixes an
assertion failure that would occur intermittently while compiling the
following IR:
define i32 @t(i32 %i) {
%cmp = icmp sle i32 0, %i
%twin1 = select i1 %cmp, i32 %i, i32 0
%cmpinv = icmp sgt i32 0, %i
%twin2 = select i1 %cmpinv, i32 0, i32 %i
%sink = add i32 %twin1, %twin2
ret i32 %sink
}
Differential Revision: https://reviews.llvm.org/D86843
MemoryLocation has been taught about memcpy.inline, which means we can
get the memory locations read and written by it. This means DSE can
handle memcpy.inline
This patch fixes pr45956 (https://bugs.llvm.org/show_bug.cgi?id=45956 ).
To minimize its impact to the quality of generated code, I suggest enabling
this only for LTO as a start (it has two JumpThreading passes registered).
This patch contains a flag that makes JumpThreading enable it.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D84940
The test in PR47457 demonstrates a situation when candidate load's pointer's SCEV
is no loger a SCEVAddRec after loop versioning. The code there assumes that it is
always a SCEVAddRec and crashes otherwise.
This patch makes sure that we do not consider candidates for which this requirement
is broken after the versioning.
Differential Revision: https://reviews.llvm.org/D87355
Reviewed By: asbirlea
Atomic stores are modeled as MemoryDef to model the fact that they may
not be reordered, depending on the ordering constraints.
Atomic stores that are monotonic or weaker do not limit re-ordering, so
we do not have to treat them as potential read clobbers.
Note that llvm/test/Transforms/DeadStoreElimination/MSSA/atomic.ll
already contains a set of negative test cases.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D87386
This is the initial part of the implementation of the C++20 likelihood
attributes. It handles the attributes in an if statement.
Differential Revision: https://reviews.llvm.org/D85091
This patch adds isGuaranteedNotToBePoison and programUndefinedIfUndefOrPoison.
isGuaranteedNotToBePoison will be used at D75808. The latter function is used at isGuaranteedNotToBePoison.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D84242
The MemorySSAWrapperPass depends on AAResultsWrapperPass and if
MemorySSA is preserved but AAResultsWrapperPass is not, this could lead
to a crash when updating the last user of the MemorySSAWrapperPass.
Alternatively AAResultsWrapperPass could be marked preserved by GVN, but
I am not sure if that would be safe. I am not sure what is required in
order to preserve AAResultsWrapperPass. At the moment, it seems like a
couple of passes that do similar transforms to GVN are preserving it.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D87137
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
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 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
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
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
Summary:
Analyses are preserved in MemCpyOptimizer.
Get analyses before running the pass and store the pointers, instead of
using lambdas and getting them every time on demand.
Reviewers: lenary, deadalnix, mehdi_amini, nikic, efriedma
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74494
Loop Idiom Recognize Pass (LIRP) attempts to transform loops with subscripted arrays
into memcpy/memset function calls. In some particular situation, this transformation
introduces negative impacts. For example: https://bugs.llvm.org/show_bug.cgi?id=47300
This patch will enable users to disable a particular part of the transformation, while
he/she can still enjoy the benefit brought about by the rest of LIRP. The default
behavior stays unchanged: no part of LIRP is disabled by default.
Reviewed By: etiotto (Ettore Tiotto)
Differential Revision: https://reviews.llvm.org/D86262
The 1st try was reverted because I missed an assert that
needed softening.
As discussed in D86798 / rG09652721 , we were potentially
returning a different result for whether an Instruction
is commutable depending on if we call the base class or
derived class method.
This requires relaxing asserts in GVN, but that pass
seems to be working otherwise.
NewGVN requires more work because it uses different
code paths for numbering binops and calls.
For an instruction in the basic block BB, SinkingPass enumerates basic blocks
dominated by BB and BB's successors. For each enumerated basic block,
SinkingPass uses `AllUsesDominatedByBlock` to check whether the basic
block dominates all of the instruction's users. This is inefficient.
Use the nearest common dominator of all users to avoid enumerating the
candidate. The nearest common dominator may be in a parent loop which is
not beneficial. In that case, find the ancestors in the dominator tree.
In the case that the instruction has no user, with this change we will
not perform unnecessary move. This causes some amdgpu test changes.
A stage-2 x86-64 clang is a byte identical with this change.
As discussed in D86798 / rG09652721 , we were potentially
returning a different result for whether an Instruction
is commutable depending on if we call the base class or
derived class method.
This requires relaxing an assert in GVN, but that pass
seems to be working otherwise.
NewGVN requires more work because it uses different
code paths for numbering binops and calls.
Handling the new min/max intrinsics is the motivation, but it
turns out that we have a bunch of other intrinsics with this
missing bit of analysis too.
The FP min/max tests show that we are intersecting FMF,
so that part should be safe too.
As noted in https://llvm.org/PR46897 , there is a commutative
property specifier for intrinsics, but no corresponding function
attribute, and so apparently no uses of that bit. We may want to
remove that next.
Follow-up patches should wire up the Instruction::isCommutative()
to this IntrinsicInst specialization. That requires updating
callers to be aware of the more general commutative property
(not just binops).
Differential Revision: https://reviews.llvm.org/D86798
This patch fixes this crash https://gcc.godbolt.org/z/Ps8d1e
And gives SROA the ability to remove assumes if it allows promoting an alloca to register
Without removing assumes when it can't promote to register.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D86570
David Sherwood