Some post-processing involved in exchanging a pair of loops has been
done separately from `processLoop`, which is a main function that does
the transformation. It's better to consolidate these processes into the
same function. This patch is a preparation of #127474.
Parameter PossiblyLoopIndependent has lost its intended purpose. This
flag is always set to true in all cases when depends() is called, hence
we want to reconsider the utility of this variable and remove it from
the function signature entirely. This is an NFC patch.
The profiling of the LLVM Test-suite reveals that a significant portion,
specifically 14,090 out of 139,323, loop nests were identified as
non-viable candidates for transformation, leading to the transform
exiting from isComputableLoopNest() without any action.
More importantly, dependence information was computed for these loop
nests before reaching the function isComputableLoopNest(), which does
not require DI and relies solely on scalar evolution (SE).
To enhance compile-time efficiency, this patch moves the call to
isComputableLoopNest() earlier in the control-flow, thereby avoiding
unnecessary dependence calculations.
The impact of this change is evident on the compile-time-tracker, with
the overall geometric mean improvement recorded at 0.11%, while the
lencode benchmark gets a more substantial benefit of 0.44%.
This improvement can be tracked in the isc-ln-exp-2 branch under my
repo.
LoopInterchange have converted `DVEntry::LE` and `DVEntry::GE` in
direction vectors to '<' and '>' respectively. This handling is
incorrect because the information about the '=' it lost. This leads to
miscompilation in some cases. To resolve this issue, convert them to '*'
instead.
Resolve#123920
As part of the "RemoveDIs" project, BasicBlock::iterator now carries a
debug-info bit that's needed when getFirstNonPHI and similar feed into
instruction insertion positions. Call-sites where that's necessary were
updated a year ago; but to ensure some type safety however, we'd like to
have all calls to getFirstNonPHI use the iterator-returning version.
This patch changes a bunch of call-sites calling getFirstNonPHI to use
getFirstNonPHIIt, which returns an iterator. All these call sites are
where it's obviously safe to fetch the iterator then dereference it. A
follow-up patch will contain less-obviously-safe changes.
We'll eventually deprecate and remove the instruction-pointer
getFirstNonPHI, but not before adding concise documentation of what
considerations are needed (very few).
---------
Co-authored-by: Stephen Tozer <Melamoto@gmail.com>
As part of the "RemoveDIs" project, BasicBlock::iterator now carries a
debug-info bit that's needed when getFirstNonPHI and similar feed into
instruction insertion positions. Call-sites where that's necessary were
updated a year ago; but to ensure some type safety however, we'd like to
have all calls to moveBefore use iterators.
This patch adds a (guaranteed dereferenceable) iterator-taking
moveBefore, and changes a bunch of call-sites where it's obviously safe
to change to use it by just calling getIterator() on an instruction
pointer. A follow-up patch will contain less-obviously-safe changes.
We'll eventually deprecate and remove the instruction-pointer
insertBefore, but not before adding concise documentation of what
considerations are needed (very few).
This patch is an extension to #115128.
After profiling LLVM test-suite, I see a lot of loop nest of depth more
than `MaxLoopNestDepth` which is 10. Early exit for them would save
compile-time as it would avoid computing DependenceInfo and CacheCost.
Please see 'bound-max-depth' branch on compile-time-tracker.
In the current state of the code, the transform computes entries for the
dependency matrix until `MaxMemInstrCount` which is 100. After 99th
entry, it terminates and thus overall wastes compile-time.
It would be nice if we can compute total number of entries upfront and
early exit if the number of entries > 100. However, computing the number
of entries is not always possible as it depends on two factors:
1. Number of load-store pairs in a loop.
2. Number of common loop levels for each of the pair.
This patch constrains the whole computation on the number of loads and
stores instructions in the loop.
In another approach, I experimented with computing 1 and constraining
the number of pairs, but that did not lead to any additional benefit in
terms of compile time. However, when other issues are fixed, I can
revisit this approach.
We are not handling 'S' scalar dependencies correctly and have at least
the following miscompiles related to that:
[LoopInterchange] incorrect handling of scalar dependencies and dependence vectors starting with ">" #54176
[LoopInterchange] Interchange breaks program correctness #46867
[LoopInterchange] Loops should not interchanged due to dependencies #47259
[LoopInterchange] Loops should not interchanged due to control flow #47401
This patch does no longer insert the "S" dependency/direction into the
dependency matrix, so a dependency is never "S". We seem to have
forgotten what the exact meaning is of this dependency type, and don't
see why it should be treated differently.
We prefer correctness over incorrect and more aggressive results. I.e.,
this prevents the miscompiles at the expense of handling less cases,
i.e. making interchange more pessimistic. However, some of the cases
that are now rejected for dependence analysis reasons, were rejected
before too but for other reasons (e.g. profitability). So at least for
the llvm regression tests, the number of regression are very reasonable.
This should be a stopgap. We would like to get interchange enabled by
default and thus prefer correctness over unsafe transforms, and later
see if we can get solve the regressions.
The entries in the dependency matrix can contain a lot of duplicates,
which is unnecessary and results in more checks that we can avoid, and
this patch adds that.
This patch bails out early if minimum depth
is not met. As it stands today, the pass computes
CacheCost before it attempts to do the transform.
This is not needed if minimum depth is not met.
This handles basic cases where depth is typically 1.
As the patch avoids unnecessary computation, it is aimed to improve
compile-time.
As outlined in my proposal of how to get rid of debug intrinsics, this
patch adds a moveBefore method that signals the caller /intends/ the order
of moved instructions is to stay the same. This semantic difference has an
effect on debug-info, as it signals whether debug-info needs to move with
instructions or not.
The patch just replaces a few calls to moveBefore with calls to
moveBeforePreserving -- and the latter just calls the former, so it's all
NFC right now. A future patch will add an implementation of
moveBeforePreserving that takes action to correctly preserve debug-info,
but that's tightly coupled with our non-instruction debug-info
representation that's still being reviewed.
[0] https://discourse.llvm.org/t/rfc-instruction-api-changes-needed-to-eliminate-debug-intrinsics-from-ir/68939
Differential Revision: https://reviews.llvm.org/D156369
SCEVExpander keeps track of all instructions it inserted. However,
it currently misses some phi nodes created during LCSSA construction.
Fix this by collecting these into another argument.
This also removes the IRBuilder argument, which was added for
essentially the same purpose, but only handles the root LCSSA nodes,
not those inserted by SSAUpdater.
This was reported as a regression on D149344, but the reduced test
case also reproduces without it.
Differential Revision: https://reviews.llvm.org/D150681
Before D135808, There would be endless loop interchange posibility (no
proper priority was there in profitability check. Any profitable check
may leads to loop-interchange). With this patch, there is no endless
interchange (priority in profitable check is defined. Order of decision
is 'Cache cost' check, 'InstrOrderCost', 'Vectorization'). Corrected the
dependency checking inside isProfitableForVectorization(), corrected the
checking of bad order loops in isProfitablePerInstrOrderCost().
Reviewed By: Meinersbur, bmahjour, #loopoptwg
Differential Revision: https://reviews.llvm.org/D135808
We don't need to explicitly forget subloops because forgetting parent
loops will automatically forget their subloops
Differential Revision: https://reviews.llvm.org/D141029
The current code of validDepInterchange() enumerates cases that are
legal for interchange. This could be simplified by checking
lexicographically order of the swapped direction matrix.
Reviewed By: congzhe, Meinersbur, bmahjour
Differential Revision: https://reviews.llvm.org/D137461
This is the bugfix to the miscompile mentioned in
https://reviews.llvm.org/D132055#3814831. The IR
that reproduced the bug is added as the test case in
this patch.
What this patch does is that, during legality phase
instead of checking the phi nodes only in `InnerLoop`
and `OuterLoop`, we check phi nodes in all subloops
of the `OuterLoop`. Suppose if the loop nest is triply
nested, and `InnerLoop` and `OuterLoop` is the middle
loop and the outermost loop respectively, we'll check
phi nodes in the innermost loop as well, in addition to
the ones in the middle and outermost loops.
Reviewed By: Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D134930
isOuterMostDepPositive()
The function isOuterMostDepPositive() is checked after negative dependence
vectors are normalized to be non-negative, so there will not be any negative
dependency ('>' as the outermost non-equal sign) after normalization. And
therefore the check in isOuterMostDepPositive() is irrelevent and redundant.
Reviewed By: congzhe
Differential Revision: https://reviews.llvm.org/D132982
This is a bugfix patch that resolves the following two bugs in loop interchange:
1. PR57148 which is an assertion error due to of loss of LCSSA form after interchange,
as referred to test1() in pr57148.ll.
2. Use before def for the outermost loop induction variables after interchange,
as referred to test2() in pr57148.ll.
The fix in this patch is that:
1. In cases where the LCSSA form is not maintained after interchange, we update the IR
to the LCSSA form again.
2. We split the phi nodes in the inner loop header into a separate basic block to avoid
the situation where use of the outer indvar appears before its def after interchange.
Previously we already did this for innermost loops, now we do it for non-innermost
loops (e.g., middle loops) as well.
Reviewed By: bmahjour, Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D132055
This patch is to resolve the bug reported and discussed in
https://reviews.llvm.org/D124926#3718761 and https://reviews.llvm.org/D124926#3719876.
The problem is that loop interchange is a loopnest pass under the new pass manager,
but the loop nest may not be constructed correctly by the loop pass manager after
running loop interchange and before running the next pass, which might cause problems
when it continues running the next pass.
The reason that the loop nest is constructed incorrectly is that the outermost
loop might have changed after interchange, and what was the original outermost
loop is not the current outermost loop anymore. Constructing the loop nest based
on the original outermost loop would generate an invalid loop nest.
The fix in this patch is that, in the loop pass manager before running each loopnest
pass, we re-cosntruct the loop nest based on the current outermost loop, if LPMUpdater
notifies the loop pass manager that the previous loop nest has been invalidated by passes
like loop interchange.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D132199
This is the 2nd patch of the two-patch series (D130188, D130189) that
fix PR56275 (https://github.com/llvm/llvm-project/issues/56275) which
is a missed opportunity for loop interchange.
As follow-up on the dependence analysis (DA) patch D130188, this patch
normalizes DA results in loop interchange, such that negative dependence
vectors queried by loop interchange are reversed to be non-negative.
Now all tests in PR56275 can get interchanged. Those tests are added
in lit test as `pr56275.ll`.
Reviewed By: kawashima-fj, bmahjour, Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D130189
This is another attempt to land this patch.
The patch proposed to use a new cost model for loop interchange,
which is obtained from loop cache analysis.
Given a loopnest, what loop cache analysis returns is a vector of
loops [loop0, loop1, loop2, ...] where loop0 should be replaced as
the outermost loop, loop1 should be placed one more level inside, and
loop2 one more level inside, etc. What loop cache analysis does is not
only more comprehensive than the current cost model, it is also a "one-shot"
query which means that we only need to query it once during the entire
loop interchange pass, which is better than the current cost model where
we query it every time we check whether it is profitable to interchange
two loops. Thus complexity is reduced, especially after D120386 where we
do more interchanges to get the globally optimal loop access pattern.
Updates made to test cases are mostly minor changes and some
corrections. One change that applies to all tests is that we added an option
`-cache-line-size=64` to the RUN lines. This is ensure that loop
cache analysis receives a valid number of cache line size for correct
analysis. Test coverage for loop interchange is not reduced.
Currently we did not completely remove the legacy cost model, but
keep it as fall-back in case the new cost model did not run successfully.
This is because currently we have some limitations in delinearization, which
sometimes makes loop cache analysis bail out. The longer term goal is to
enhance delinearization and eventually remove the legacy cost model
compeletely.
Reviewed By: bmahjour, #loopoptwg
Differential Revision: https://reviews.llvm.org/D124926
llvm/lib/Analysis/LoopCacheAnalysis.cpp:702:30: runtime error: signed
integer overflow: 6148914691236517209 * 100 cannot be represented in
type 'long'
https://lab.llvm.org/buildbot/#/builders/5/builds/25185
This reverts commit 1b24fe34b06cd9f2337313f513a8b19f9a37c5de.
This is the second attempt to land this patch.
The patch proposed to use a new cost model for loop interchange,
which is obtained from loop cache analysis.
Given a loopnest, what loop cache analysis returns is a vector of
loops [loop0, loop1, loop2, ...] where loop0 should be replaced as the
outermost loop, loop1 should be placed one more level inside, and loop2
one more level inside, etc. What loop cache analysis does is not only more
comprehensive than the current cost model, it is also a "one-shot" query
which means that we only need to query it once during the entire loop
interchange pass, which is better than the current cost model where we
query it every time we check whether it is profitable to interchange two
loops. Thus complexity is reduced, especially after D120386 where we do
more interchanges to get the globally optimal loop access pattern.
Updates made to test cases are mostly minor changes and some corrections.
One change that applies to all tests is that we added an option
`-cache-line-size=64` to the RUN lines. This is ensure that loop cache
analysis receives a valid number of cache line size for correct analysis.
Test coverage for loop interchange is not reduced.
Currently we did not completely remove the legacy cost model, but keep it
as fall-back in case the new cost model did not run successfully. This is
because currently we have some limitations in delinearization, which sometimes
makes loop cache analysis bail out. The longer term goal is to enhance
delinearization and eventually remove the legacy cost model compeletely.
Reviewed By: bmahjour, #loopoptwg
Differential Revision: https://reviews.llvm.org/D124926
This patch proposed to use a new cost model for loop interchange, which
is obtained from loop cache analysis.
Given a loopnest, what loop cache analysis returns is a vector of loops
[loop0, loop1, loop2, ...] where loop0 should be replaced as the outermost
loop, loop1 should be placed one more level inside, and loop2 one more level
inside, etc. What loop cache analysis does is not only more comprehensive than
the current cost model, it is also a "one-shot" query which means that we only
need to query it once during the entire loop interchange pass, which is better
than the current cost model where we query it every time we check whether it is
profitable to interchange two loops. Thus complexity is reduced, especially after
D120386 where we do more interchanges to get the globally optimal loop access pattern.
Updates made to test cases are mostly minor changes and some corrections.
Test coverage for loop interchange is not reduced.
Currently we did not completely remove the legacy cost model, but keep it as
fall-back in case the new cost model did not run successfully. This is because
currently we have some limitations in delinearization, which sometimes makes
loop cache analysis bail out. The longer term goal is to enhance delinearization
and eventually remove the legacy cost model compeletely.
Reviewed By: bmahjour, #loopoptwg
Differential Revision: https://reviews.llvm.org/D124926
Motivated by pr43326 (https://bugs.llvm.org/show_bug.cgi?id=43326), where a slightly
modified case is as follows.
void f(int e[10][10][10], int f[10][10][10]) {
for (int a = 0; a < 10; a++)
for (int b = 0; b < 10; b++)
for (int c = 0; c < 10; c++)
f[c][b][a] = e[c][b][a];
}
The ideal optimal access pattern after running interchange is supposed to be the following
void f(int e[10][10][10], int f[10][10][10]) {
for (int c = 0; c < 10; c++)
for (int b = 0; b < 10; b++)
for (int a = 0; a < 10; a++)
f[c][b][a] = e[c][b][a];
}
Currently loop interchange is limited to picking up the innermost loop and finding an order
that is locally optimal for it. However, the pass failed to produce the globally optimal
loop access order. For more complex examples what we get could be quite far from the
globally optimal ordering.
What is proposed in this patch is to do a "bubble-sort" fashion when doing interchange.
By comparing neighbors in `LoopList` in each iteration, we would be able to move each loop
onto a most appropriate place, hence this is an approach that tries to achieve the
globally optimal ordering.
The motivating example above is added as a test case.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D120386
This patch is motivated by pr48057 where an output dependency is not detected
since loop interchange did not check a store instruction with itself.
Fixed that deficiency.
Reviewed By: bmahjour, Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D118102
Enabled loop interchange support for floating point reductions
if it is allowed to reorder floating point operations.
Previously when we encouter a floating point PHI node in the
outer loop exit block, we bailed out since we could not detect
floating point reductions in the early days. Now we remove this
limiation since we are able to detect floating point reductions.
Reviewed By: #loopoptwg, Meinersbur
Differential Revision: https://reviews.llvm.org/D117450
Currently loop interchange only supports loops with one inner loop
induction variable. This patch adds support for transformation with
more than one inner loop induction variables. The induction PHIs and
induction increment instructions are moved/duplicated properly to the
new outer header and the new outer latch, respectively.
Reviewed By: bmahjour
Differential Revision: https://reviews.llvm.org/D114917
This patch enables loop interchange with multiple outer loop
induction variables, and hence removes the limitation that only
a single outer loop induction variable is supported. In fact, it
turns out that the current pass already trivially supports multiple
outer indvars, which is the result of a previous patch
`https://reviews.llvm.org/D102743`. Therefore, this patch removed that
limitation and provides test cases for multiple outer indvars.
Reviewed By: bmahjour
Differential Revision: https://reviews.llvm.org/D114916
There was a limitation in legality that in the original inner loop latch,
no instruction was allowed between the induction variable increment
and the branch instruction. This is because we used to split the
inner latch at the induction variable increment instruction. Since
now we have split at the inner latch branch instruction and have
properly duplicated instructions over to the split block, we remove
this limitation.
Please refer to the test case updates to see how we now interchange
loops where instructions exist between the induction variable
increment and the branch instruction.
Reviewed By: bmahjour
Differential Revision: https://reviews.llvm.org/D115238
Patch that removed the use of this variable was reverted in
8ade3d43a3e48eb739c9db2f38b618fa213f0546
This reverts commit 3988a06d86e1a14dfd5f5fdae84ddbf928e85dab.
There was a limitation in legality that in the original inner loop latch,
no instruction was allowed between the induction variable increment
and the branch instruction. This is because we used to split the
inner latch at the induction variable increment instruction. Since
now we have split at the inner latch branch instruction and have
properly duplicated instructions over to the split block, we remove
this limitation.
Please refer to the test case updates to see how we now interchange
loops where instructions exist between the induction variable increment
and the branch instruction.
Reviewed By: bmahjour
Differential Revision: https://reviews.llvm.org/D115238
The naming has come up as a source of confusion in several recent reviews. onlyWritesMemory is consist with onlyReadsMemory which we use for the corresponding readonly case as well.
