Skip to content

[DFAJumpThreading] Add an early exit heuristic for unpredictable values #85015

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Mar 16, 2024
Merged

[DFAJumpThreading] Add an early exit heuristic for unpredictable values #85015

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4d90635
[DFAJumpThreading] Add an early exit heuristic for unpredictable values
UsmanNadeem Mar 13, 2024
968137a
check for subloops as well
UsmanNadeem Mar 15, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 31 additions & 10 deletions llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,6 +96,11 @@ static cl::opt<bool>
cl::desc("View the CFG before DFA Jump Threading"),
cl::Hidden, cl::init(false));

static cl::opt<bool> EarlyExitHeuristic(
"dfa-early-exit-heuristic",
cl::desc("Exit early if an unpredictable value come from the same loop"),
cl::Hidden, cl::init(true));

static cl::opt<unsigned> MaxPathLength(
"dfa-max-path-length",
cl::desc("Max number of blocks searched to find a threading path"),
Expand Down Expand Up @@ -405,7 +410,7 @@ struct MainSwitch {
///
/// Also, collect select instructions to unfold.
bool isCandidate(const SwitchInst *SI) {
std::deque<Value *> Q;
std::deque<std::pair<Value *, BasicBlock *>> Q;
SmallSet<Value *, 16> SeenValues;
SelectInsts.clear();

Expand All @@ -415,25 +420,28 @@ struct MainSwitch {
return false;

// The switch must be in a loop.
if (!LI->getLoopFor(SI->getParent()))
const Loop *L = LI->getLoopFor(SI->getParent());
if (!L)
return false;

addToQueue(SICond, Q, SeenValues);
addToQueue(SICond, nullptr, Q, SeenValues);

while (!Q.empty()) {
Value *Current = Q.front();
Value *Current = Q.front().first;
BasicBlock *CurrentIncomingBB = Q.front().second;
Q.pop_front();

if (auto *Phi = dyn_cast<PHINode>(Current)) {
for (Value *Incoming : Phi->incoming_values()) {
addToQueue(Incoming, Q, SeenValues);
for (BasicBlock *IncomingBB : Phi->blocks()) {
Value *Incoming = Phi->getIncomingValueForBlock(IncomingBB);
addToQueue(Incoming, IncomingBB, Q, SeenValues);
}
LLVM_DEBUG(dbgs() << "\tphi: " << *Phi << "\n");
} else if (SelectInst *SelI = dyn_cast<SelectInst>(Current)) {
if (!isValidSelectInst(SelI))
return false;
addToQueue(SelI->getTrueValue(), Q, SeenValues);
addToQueue(SelI->getFalseValue(), Q, SeenValues);
addToQueue(SelI->getTrueValue(), CurrentIncomingBB, Q, SeenValues);
addToQueue(SelI->getFalseValue(), CurrentIncomingBB, Q, SeenValues);
LLVM_DEBUG(dbgs() << "\tselect: " << *SelI << "\n");
if (auto *SelIUse = dyn_cast<PHINode>(SelI->user_back()))
SelectInsts.push_back(SelectInstToUnfold(SelI, SelIUse));
Expand All @@ -446,18 +454,31 @@ struct MainSwitch {
// initial switch values that can be ignored (they will hit the
// unthreaded switch) but this assumption will get checked later after
// paths have been enumerated (in function getStateDefMap).

// If the unpredictable value comes from the same inner loop it is
// likely that it will also be on the enumerated paths, causing us to
// exit after we have enumerated all the paths. This heuristic save
// compile time because a search for all the paths can become expensive.
if (EarlyExitHeuristic &&
L->contains(LI->getLoopFor(CurrentIncomingBB))) {
LLVM_DEBUG(dbgs()
<< "\tExiting early due to unpredictability heuristic.\n");
return false;
}

continue;
}
}

return true;
}

void addToQueue(Value *Val, std::deque<Value *> &Q,
void addToQueue(Value *Val, BasicBlock *BB,
std::deque<std::pair<Value *, BasicBlock *>> &Q,
SmallSet<Value *, 16> &SeenValues) {
if (SeenValues.contains(Val))
return;
Q.push_back(Val);
Q.push_back({Val, BB});
SeenValues.insert(Val);
}

Expand Down
2 changes: 1 addition & 1 deletion llvm/test/Transforms/DFAJumpThreading/dfa-unfold-select.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -S -passes=dfa-jump-threading %s | FileCheck %s
; RUN: opt -S -passes=dfa-jump-threading -dfa-early-exit-heuristic=false %s | FileCheck %s

; These tests check if selects are unfolded properly for jump threading
; opportunities. There are three different patterns to consider:
Expand Down
124 changes: 124 additions & 0 deletions llvm/test/Transforms/DFAJumpThreading/unpredictable-heuristic.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,124 @@
; REQUIRES: asserts
; RUN: opt -S -passes=dfa-jump-threading %s -debug-only=dfa-jump-threading 2>&1 | FileCheck %s

; CHECK-COUNT-3: Exiting early due to unpredictability heuristic.

@.str.1 = private unnamed_addr constant [3 x i8] c"10\00", align 1
@.str.2 = private unnamed_addr constant [3 x i8] c"30\00", align 1
@.str.3 = private unnamed_addr constant [3 x i8] c"20\00", align 1
@.str.4 = private unnamed_addr constant [3 x i8] c"40\00", align 1

define void @test1(i32 noundef %num, i32 noundef %num2) {
entry:
br label %while.body

while.body: ; preds = %entry, %sw.epilog
%num.addr.0 = phi i32 [ %num, %entry ], [ %num.addr.1, %sw.epilog ]
switch i32 %num.addr.0, label %sw.default [
i32 10, label %sw.bb
i32 30, label %sw.bb1
i32 20, label %sw.bb2
i32 40, label %sw.bb3
]

sw.bb: ; preds = %while.body
%call.i = tail call i32 @bar(ptr noundef nonnull @.str.1)
br label %sw.epilog

sw.bb1: ; preds = %while.body
%call.i4 = tail call i32 @bar(ptr noundef nonnull @.str.2)
br label %sw.epilog

sw.bb2: ; preds = %while.body
%call.i5 = tail call i32 @bar(ptr noundef nonnull @.str.3)
br label %sw.epilog

sw.bb3: ; preds = %while.body
%call.i6 = tail call i32 @bar(ptr noundef nonnull @.str.4)
%call = tail call noundef i32 @foo()
%add = add nsw i32 %call, %num2
br label %sw.epilog

sw.default: ; preds = %while.body
ret void

sw.epilog: ; preds = %sw.bb3, %sw.bb2, %sw.bb1, %sw.bb
%num.addr.1 = phi i32 [ %add, %sw.bb3 ], [ 40, %sw.bb2 ], [ 20, %sw.bb1 ], [ 30, %sw.bb ]
br label %while.body
}


define void @test2(i32 noundef %num, i32 noundef %num2) {
entry:
br label %while.body

while.body: ; preds = %entry, %sw.epilog
%num.addr.0 = phi i32 [ %num, %entry ], [ %num.addr.1, %sw.epilog ]
switch i32 %num.addr.0, label %sw.default [
i32 10, label %sw.epilog
i32 30, label %sw.bb1
i32 20, label %sw.bb2
i32 40, label %sw.bb3
]

sw.bb1: ; preds = %while.body
br label %sw.epilog

sw.bb2: ; preds = %while.body
br label %sw.epilog

sw.bb3: ; preds = %while.body
br label %sw.epilog

sw.default: ; preds = %while.body
ret void

sw.epilog: ; preds = %while.body, %sw.bb3, %sw.bb2, %sw.bb1
%.str.4.sink = phi ptr [ @.str.4, %sw.bb3 ], [ @.str.3, %sw.bb2 ], [ @.str.2, %sw.bb1 ], [ @.str.1, %while.body ]
%num.addr.1 = phi i32 [ %num2, %sw.bb3 ], [ 40, %sw.bb2 ], [ 20, %sw.bb1 ], [ 30, %while.body ]
%call.i6 = tail call i32 @bar(ptr noundef nonnull %.str.4.sink)
br label %while.body
}


define void @test3(i32 noundef %num, i32 noundef %num2) {
entry:
%add = add nsw i32 %num2, 40
br label %while.body

while.body: ; preds = %entry, %sw.epilog
%num.addr.0 = phi i32 [ %num, %entry ], [ %num.addr.1, %sw.epilog ]
switch i32 %num.addr.0, label %sw.default [
i32 10, label %sw.bb
i32 30, label %sw.bb1
i32 20, label %sw.bb2
i32 40, label %sw.bb3
]

sw.bb: ; preds = %while.body
%call.i = tail call i32 @bar(ptr noundef nonnull @.str.1)
br label %sw.epilog

sw.bb1: ; preds = %while.body
%call.i5 = tail call i32 @bar(ptr noundef nonnull @.str.2)
br label %sw.epilog

sw.bb2: ; preds = %while.body
%call.i6 = tail call i32 @bar(ptr noundef nonnull @.str.3)
br label %sw.epilog

sw.bb3: ; preds = %while.body
%call.i7 = tail call i32 @bar(ptr noundef nonnull @.str.4)
br label %sw.epilog

sw.default: ; preds = %while.body
ret void

sw.epilog: ; preds = %sw.bb3, %sw.bb2, %sw.bb1, %sw.bb
%num.addr.1 = phi i32 [ %add, %sw.bb3 ], [ 40, %sw.bb2 ], [ 20, %sw.bb1 ], [ 30, %sw.bb ]
br label %while.body
}


declare noundef i32 @foo()
declare noundef i32 @bar(ptr nocapture noundef readonly)