For every threadable path `B1 -> B2 -> ... -> Bn`, we need to insert phi
nodes into every unduplicated successor of `Bi` if there are outer uses
of duplicated definitions in `B_i`. To prevent the booming of phi nodes,
this patch adds a threshold for the maximum number of unduplicated
successors that may contain outer uses. This threshold makes sense
especially when multi-target branches like switch/indirectbr/callbr are
duplicated.
Note that the O3 statistics in llvm-test-suite are not influenced.
The edge `StartBlock -> EndBlock` already exists before unfolding.
The instructions for `applyUpdates()` say that you are supposed not
to insert an existing edge.
Fixes issues reported by @mikaelholmen in
https://github.com/llvm/llvm-project/pull/162802
Note that the test coverage misses the dominator tree verification. This
patch controls verification by option, instead of using the
EXPENSIVE_CHECKS macro.
DFAJumpThreading determines the switch destination of a threadable path
based on its next state and cannot reuse cloned blocks if their
destinations differ.
However, different states may lead to the same destination. This patch
unifies equivalent states, thereby avoiding redundant duplication of
cloned blocks.
Along the lines of [#161240](vscode-file://vscode-app/Applications/Visual%20Studio%20Code.app/Contents/Resources/app/out/vs/code/electron-browser/workbench/workbench.html "Open Issue or Pull Request #161240 on GitHub")
Duplicating blocks of threaded paths may cause a significant regression
in IR size and slow down compile-time in later optimizations. This patch
adds a coarse constraint on the number of duplicated instructions.
Fixes#160250
We previously assumed the select to unfold is defined in the incoming
block of phi user, as `isValidSelectInst` filters other cases at the
initial stage. However, the selects not defined in the incoming block
may occur after unfolding the arms of the unfolded select.
This patch sinks the select into the incoming block of the phi user and
unfolds it at the incoming block.
The limit 'dfa-max-num-paths' that is used to control number of
enumerated paths was not checked against inside getPathsFromStateDefMap.
It may lead to large memory consumption for complex enough switch
statements.
Reland llvm/llvm-project#145482
We just replaced SmallSet<T *, N> with SmallPtrSet<T *, N>, bypassing
the redirection found in SmallSet.h. With that, we no longer need to
include SmallSet.h in many files.
This patch replaces SmallSet<T *, N> with SmallPtrSet<T *, N>. Note
that SmallSet.h "redirects" SmallSet to SmallPtrSet for pointer
element types:
template <typename PointeeType, unsigned N>
class SmallSet<PointeeType*, N> : public SmallPtrSet<PointeeType*, N>
{};
We only have 140 instances that rely on this "redirection", with the
vast majority of them under llvm/. Since relying on the redirection
doesn't improve readability, this patch replaces SmallSet with
SmallPtrSet for pointer element types.
The limit 'dfa-max-num-paths' that is used to control number of
enumerated paths was not checked against inside getPathsFromStateDefMap.
It may lead to large memory consumption for complex enough switch
statements.
We can use *Set::insert_range to collapse:
for (auto Elem : Range)
Set.insert(E);
down to:
Set.insert_range(Range);
In some cases, we can further fold that into the set declaration.
After #96127 landed, mshockwave reported that the pass was no longer
threading SPEC2006/perlbench.
After 96127 we started bailing out in `getStateDefMap` and rejecting the
transformation because one of the unpredictable values was coming from
inside the loop. There was no fundamental change in that function except
that we started calling `Loop->contains(IncomingBB)` instead of
`LoopBBs.count(IncomingBB)`. After some analysis I came to the
conclusion that even before 96127 we would reject the transformation if
we provided large enough limits on the path traversal (large enough so
that LoopBBs contained blocks corresponding to that unpredictable
value).
In this patch I changed `getStateDefMap` to not terminate early on
finding an unpredictable value, this is because
`getPathsFromStateDefMap`, later, actually has checks to ensure that the
final list of paths only have predictable values. As a result we can now
partially thread functions like `negative6` in the tests that have some
predictable paths.
This patch does not really have any compile-time impact on the test
suite without `-dfa-early-exit-heuristic=false` (early exit is enabled
by default).
Change-Id: Ie1633b370ed4a0eda8dea52650b40f6f66ef49a3
…ect successor
Previously the code assumed that the select instruction is defined in a
block that is a direct predecessor of the block where the PHINode uses
it. So, we were hitting an assertion when we tried to access the def
block as an incoming block for the user phi node.
This patch handles that case by using the correct end block and creating
a new phi node that aggregates both the values of the select in that end
block, and then using that new unfolded phi to overwrite the original
user phi node.
Fixes#106083
Change-Id: Ie471994cca232318f74a6e6438efa21e561c2dc0
I tried to add a limit to number of blocks visited in the paths()
function but even with a very high limit the transformation coverage was
being reduced.
After looking at the code it seemed that the function was trying to
create paths of the form
`SwitchBB...DeterminatorBB...SwitchPredecessor`. This is inefficient
because a lot of nodes in those paths (nodes before DeterminatorBB)
would be irrelevant to the optimization. We only care about paths of the
form `DeterminatorBB_Pred DeterminatorBB...SwitchBB`. This weeds out a
lot of visited nodes.
In this patch I have added a hard limit to the number of nodes visited
and changed the algorithm for path calculation. Primarily I am
traversing the use-def chain for the PHI nodes that define the state. If
we have a hole in the use-def chain (no immediate predecessors) then I
call the paths() function.
I also had to the change the select instruction unfolding code to insert
redundant one input PHIs to allow the use of the use-def chain in
calculating the paths.
The test suite coverage with this patch (including a limit on nodes
visited) is as follows:
Geomean diff:
dfa-jump-threading.NumTransforms: +13.4%
dfa-jump-threading.NumCloned: +34.1%
dfa-jump-threading.NumPaths: -80.7%
Compile time effect vs baseline (pass enabled by default) is mostly
positive:
https://llvm-compile-time-tracker.com/compare.php?from=ad8705fda25f64dcfeb6264ac4d6bac36bee91ab&to=5a3af6ce7e852f0736f706b4a8663efad5bce6ea&stat=instructions:u
Change-Id: I0fba9e0f8aa079706f633089a8ccd4ecf57547ed
- There is no restriction on a loop with controlled convergent
operations when
the relevant tokens are defined and used within the loop.
- When a token defined outside a loop is used inside (also called a loop
convergence heart), unrolling is allowed only in the absence of
remainder or
runtime checks.
- When a token defined inside a loop is used outside, such a loop is
said to be
"extended". This loop can only be unrolled by also duplicating the
extended part
lying outside the loop. Such unrolling is disabled for now.
- Clean up loop hearts: When unrolling a loop with a heart, duplicating
the
heart will introduce multiple static uses of a convergence control token
in a
cycle that does not contain its definition. This violates the static
rules for
tokens, and needs to be cleaned up into a single occurrence of the
intrinsic.
- Spell out the initializer for UnrollLoopOptions to improve
readability.
Original implementation [D85605] by Nicolai Haehnle
<nicolai.haehnle@amd.com>.
Right now the algorithm does not exit on unpredictable values. It
waits until all the paths have been enumerated to see if any of
those paths have that value. Waiting this late leads to a lot of
wasteful computation and higher compile time.
In this patch I have added a heuristic that checks if the value
comes from the same inner loops as the switch, if so, then it is
likely that the value will also be seen on a threadable path and
the code in `getStateDefMap()` return an empty map.
I tested this on the llvm test suite and the only change in the
number of threaded switches was in 7zip (before 23, after 18).
In all of those cases the current algorithm was partially threading
the loop because it was hitting a limit on the number of paths to
be explored. On increasing this limit even the current algorithm
finds paths where the unpredictable value is seen.
Compile time(with pass enabled by default and this patch):
https://llvm-compile-time-tracker.com/compare.php?from=8c5e9cf737138aba22a4a8f64ef2c5efc80dd7f9&to=42c75d888058b35c6d15901b34e36251d8f766b9&stat=instructions:u
Fixes the buildbot failure in
https://github.com/llvm/llvm-project/pull/78134#issuecomment-1892195197
When we meet the path with single `determinator`, the determinator
actually takes itself as a predecessor. Thus, we need to let `Prev` be
the determinator when `PathBBs` has only one element.
Fixes#56882.
Fixes#60254.
When switch has only one successor, it make no sense to thread it. And
computing the cost of it brings div-by-zero exception. We prevent it in
this patch.
Fixes#64860.
When a select instruction comes in by PHINode, the phi's incoming block
for it can flow indirectly past other BasicBlock into it. In this case,
we cannot unfold select to the phi's BB.
DFAJumpThreading
JumpThreading
LibCallsShrink
LoopVectorize
SLPVectorizer
DeadStoreElimination
AggressiveDCE
CorrelatedValuePropagation
IndVarSimplify
These are part of the optimization pipeline, of which the legacy version is deprecated and being removed.
Per the documentation in Support/InstructionCost.h, the purpose of an invalid cost is so that clients can change behavior on impossible to cost inputs. CodeMetrics was instead asserting that invalid costs never occurred.
On a target with an incomplete cost model - e.g. RISCV - this means that transformations would crash on (falsely) invalid constructs - e.g. scalable vectors. While we certainly should improve the cost model - and I plan to do so in the near future - we also shouldn't be crashing. This violates the explicitly stated purpose of an invalid InstructionCost.
I updated all of the "easy" consumers where bailouts were locally obvious. I plan to follow up with loop unroll in a following change.
Differential Revision: https://reviews.llvm.org/D127131
