[SimpleLoopUnswitch] Fix exponential unswitch #66883
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-llvm-transforms ChangesWhen unswitching via invariant condition injection, we currently mark the condition in the old loop, so that it does not get unswitched again. However, if there are multiple branches for which conditions can be injected, then we can do that for both the old and new loops. This means that the number of unswitches increases exponentially. Change the handling to be more similar to partial unswitching, where we instead mark the whole loop, rather than a single condition. This means that we will only generate a linear number of loops. TBH I think even that is still highly undesirable, and we should probably be unswitching all candidates at the same time, so that we end up with only two loops. But at least this mitigates the worst case. The test case is a reduced variant that generates 1700 lines of IR without this patch and 290 with it. Fixes #66868. Patch is 35.43 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/66883.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp b/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp
index cad026ffbbd6b74..8ac974bdfecf84a 100644
--- a/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp
+++ b/llvm/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp
@@ -2127,9 +2127,10 @@ static void unswitchNontrivialInvariants(
Loop &L, Instruction &TI, ArrayRef<Value *> Invariants,
IVConditionInfo &PartialIVInfo, DominatorTree &DT, LoopInfo &LI,
AssumptionCache &AC,
- function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB,
+ function_ref<void(bool, bool, bool, ArrayRef<Loop *>)> UnswitchCB,
ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
- function_ref<void(Loop &, StringRef)> DestroyLoopCB, bool InsertFreeze) {
+ function_ref<void(Loop &, StringRef)> DestroyLoopCB, bool InsertFreeze,
+ bool InjectedCondition) {
auto *ParentBB = TI.getParent();
BranchInst *BI = dyn_cast<BranchInst>(&TI);
SwitchInst *SI = BI ? nullptr : cast<SwitchInst>(&TI);
@@ -2582,7 +2583,7 @@ static void unswitchNontrivialInvariants(
for (Loop *UpdatedL : llvm::concat<Loop *>(NonChildClonedLoops, HoistedLoops))
if (UpdatedL->getParentLoop() == ParentL)
SibLoops.push_back(UpdatedL);
- UnswitchCB(IsStillLoop, PartiallyInvariant, SibLoops);
+ UnswitchCB(IsStillLoop, PartiallyInvariant, InjectedCondition, SibLoops);
if (MSSAU && VerifyMemorySSA)
MSSAU->getMemorySSA()->verifyMemorySSA();
@@ -2980,13 +2981,6 @@ static bool shouldTryInjectInvariantCondition(
/// the metadata.
bool shouldTryInjectBasingOnMetadata(const BranchInst *BI,
const BasicBlock *TakenSucc) {
- // Skip branches that have already been unswithed this way. After successful
- // unswitching of injected condition, we will still have a copy of this loop
- // which looks exactly the same as original one. To prevent the 2nd attempt
- // of unswitching it in the same pass, mark this branch as "nothing to do
- // here".
- if (BI->hasMetadata("llvm.invariant.condition.injection.disabled"))
- return false;
SmallVector<uint32_t> Weights;
if (!extractBranchWeights(*BI, Weights))
return false;
@@ -3069,13 +3063,9 @@ injectPendingInvariantConditions(NonTrivialUnswitchCandidate Candidate, Loop &L,
Builder.CreateCondBr(InjectedCond, InLoopSucc, CheckBlock);
Builder.SetInsertPoint(CheckBlock);
- auto *NewTerm = Builder.CreateCondBr(TI->getCondition(), TI->getSuccessor(0),
- TI->getSuccessor(1));
-
+ Builder.CreateCondBr(TI->getCondition(), TI->getSuccessor(0),
+ TI->getSuccessor(1));
TI->eraseFromParent();
- // Prevent infinite unswitching.
- NewTerm->setMetadata("llvm.invariant.condition.injection.disabled",
- MDNode::get(BB->getContext(), {}));
// Fixup phis.
for (auto &I : *InLoopSucc) {
@@ -3443,7 +3433,7 @@ static bool shouldInsertFreeze(Loop &L, Instruction &TI, DominatorTree &DT,
static bool unswitchBestCondition(
Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC,
AAResults &AA, TargetTransformInfo &TTI,
- function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB,
+ function_ref<void(bool, bool, bool, ArrayRef<Loop *>)> UnswitchCB,
ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
function_ref<void(Loop &, StringRef)> DestroyLoopCB) {
// Collect all invariant conditions within this loop (as opposed to an inner
@@ -3453,9 +3443,10 @@ static bool unswitchBestCondition(
Instruction *PartialIVCondBranch = nullptr;
collectUnswitchCandidates(UnswitchCandidates, PartialIVInfo,
PartialIVCondBranch, L, LI, AA, MSSAU);
- collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,
- PartialIVCondBranch, L, DT, LI, AA,
- MSSAU);
+ if (!findOptionMDForLoop(&L, "llvm.loop.unswitch.injection.disable"))
+ collectUnswitchCandidatesWithInjections(UnswitchCandidates, PartialIVInfo,
+ PartialIVCondBranch, L, DT, LI, AA,
+ MSSAU);
// If we didn't find any candidates, we're done.
if (UnswitchCandidates.empty())
return false;
@@ -3476,8 +3467,11 @@ static bool unswitchBestCondition(
return false;
}
- if (Best.hasPendingInjection())
+ bool InjectedCondition = false;
+ if (Best.hasPendingInjection()) {
Best = injectPendingInvariantConditions(Best, L, DT, LI, AC, MSSAU);
+ InjectedCondition = true;
+ }
assert(!Best.hasPendingInjection() &&
"All injections should have been done by now!");
@@ -3505,7 +3499,7 @@ static bool unswitchBestCondition(
<< ") terminator: " << *Best.TI << "\n");
unswitchNontrivialInvariants(L, *Best.TI, Best.Invariants, PartialIVInfo, DT,
LI, AC, UnswitchCB, SE, MSSAU, DestroyLoopCB,
- InsertFreeze);
+ InsertFreeze, InjectedCondition);
return true;
}
@@ -3534,7 +3528,7 @@ static bool
unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC,
AAResults &AA, TargetTransformInfo &TTI, bool Trivial,
bool NonTrivial,
- function_ref<void(bool, bool, ArrayRef<Loop *>)> UnswitchCB,
+ function_ref<void(bool, bool, bool, ArrayRef<Loop *>)> UnswitchCB,
ScalarEvolution *SE, MemorySSAUpdater *MSSAU,
ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI,
function_ref<void(Loop &, StringRef)> DestroyLoopCB) {
@@ -3549,7 +3543,8 @@ unswitchLoop(Loop &L, DominatorTree &DT, LoopInfo &LI, AssumptionCache &AC,
if (Trivial && unswitchAllTrivialConditions(L, DT, LI, SE, MSSAU)) {
// If we unswitched successfully we will want to clean up the loop before
// processing it further so just mark it as unswitched and return.
- UnswitchCB(/*CurrentLoopValid*/ true, false, {});
+ UnswitchCB(/*CurrentLoopValid*/ true, /*PartiallyInvariant*/ false,
+ /*InjectedCondition*/ false, {});
return true;
}
@@ -3645,6 +3640,7 @@ PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
auto UnswitchCB = [&L, &U, &LoopName](bool CurrentLoopValid,
bool PartiallyInvariant,
+ bool InjectedCondition,
ArrayRef<Loop *> NewLoops) {
// If we did a non-trivial unswitch, we have added new (cloned) loops.
if (!NewLoops.empty())
@@ -3664,6 +3660,16 @@ PreservedAnalyses SimpleLoopUnswitchPass::run(Loop &L, LoopAnalysisManager &AM,
Context, L.getLoopID(), {"llvm.loop.unswitch.partial"},
{DisableUnswitchMD});
L.setLoopID(NewLoopID);
+ } else if (InjectedCondition) {
+ // Do the same for injection of invariant conditions.
+ auto &Context = L.getHeader()->getContext();
+ MDNode *DisableUnswitchMD = MDNode::get(
+ Context,
+ MDString::get(Context, "llvm.loop.unswitch.injection.disable"));
+ MDNode *NewLoopID = makePostTransformationMetadata(
+ Context, L.getLoopID(), {"llvm.loop.unswitch.injection"},
+ {DisableUnswitchMD});
+ L.setLoopID(NewLoopID);
} else
U.revisitCurrentLoop();
} else
@@ -3756,6 +3762,7 @@ bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
auto *SE = SEWP ? &SEWP->getSE() : nullptr;
auto UnswitchCB = [&L, &LPM](bool CurrentLoopValid, bool PartiallyInvariant,
+ bool InjectedCondition,
ArrayRef<Loop *> NewLoops) {
// If we did a non-trivial unswitch, we have added new (cloned) loops.
for (auto *NewL : NewLoops)
@@ -3768,7 +3775,7 @@ bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
// If the current loop has been unswitched using a partially invariant
// condition, we should not re-add the current loop to avoid unswitching
// on the same condition again.
- if (!PartiallyInvariant)
+ if (!PartiallyInvariant && !InjectedCondition)
LPM.addLoop(*L);
} else
LPM.markLoopAsDeleted(*L);
diff --git a/llvm/test/Transforms/SimpleLoopUnswitch/inject-invariant-conditions-exponential.ll b/llvm/test/Transforms/SimpleLoopUnswitch/inject-invariant-conditions-exponential.ll
new file mode 100644
index 000000000000000..bd88b4030d675e6
--- /dev/null
+++ b/llvm/test/Transforms/SimpleLoopUnswitch/inject-invariant-conditions-exponential.ll
@@ -0,0 +1,260 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
+; RUN: opt -S -passes='simple-loop-unswitch<nontrivial>' < %s | FileCheck %s
+
+; Make sure invariant condition injection does not result in exponential
+; size increase.
+
+; FIXME: It probably shouldn't result in linear size increase either.
+
+define void @ham(i64 %arg) {
+; CHECK-LABEL: define void @ham(
+; CHECK-SAME: i64 [[ARG:%.*]]) {
+; CHECK-NEXT: bb:
+; CHECK-NEXT: [[INJECTED_COND:%.*]] = icmp ule i64 [[ARG]], [[ARG]]
+; CHECK-NEXT: [[INJECTED_COND_FR:%.*]] = freeze i1 [[INJECTED_COND]]
+; CHECK-NEXT: br i1 [[INJECTED_COND_FR]], label [[BB_SPLIT_US:%.*]], label [[BB_SPLIT:%.*]]
+; CHECK: bb.split.us:
+; CHECK-NEXT: [[INJECTED_COND1:%.*]] = icmp ule i64 [[ARG]], [[ARG]]
+; CHECK-NEXT: [[INJECTED_COND1_FR:%.*]] = freeze i1 [[INJECTED_COND1]]
+; CHECK-NEXT: br i1 [[INJECTED_COND1_FR]], label [[BB_SPLIT_US_SPLIT_US:%.*]], label [[BB_SPLIT_US_SPLIT:%.*]]
+; CHECK: bb.split.us.split.us:
+; CHECK-NEXT: [[INJECTED_COND2:%.*]] = icmp ule i64 [[ARG]], [[ARG]]
+; CHECK-NEXT: [[INJECTED_COND2_FR:%.*]] = freeze i1 [[INJECTED_COND2]]
+; CHECK-NEXT: br i1 [[INJECTED_COND2_FR]], label [[BB_SPLIT_US_SPLIT_US_SPLIT_US:%.*]], label [[BB_SPLIT_US_SPLIT_US_SPLIT:%.*]]
+; CHECK: bb.split.us.split.us.split.us:
+; CHECK-NEXT: [[INJECTED_COND3:%.*]] = icmp ule i64 [[ARG]], [[ARG]]
+; CHECK-NEXT: [[INJECTED_COND3_FR:%.*]] = freeze i1 [[INJECTED_COND3]]
+; CHECK-NEXT: br i1 [[INJECTED_COND3_FR]], label [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US:%.*]], label [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT:%.*]]
+; CHECK: bb.split.us.split.us.split.us.split.us:
+; CHECK-NEXT: [[INJECTED_COND4:%.*]] = icmp ule i64 [[ARG]], [[ARG]]
+; CHECK-NEXT: [[INJECTED_COND4_FR:%.*]] = freeze i1 [[INJECTED_COND4]]
+; CHECK-NEXT: br i1 [[INJECTED_COND4_FR]], label [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US:%.*]], label [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT:%.*]]
+; CHECK: bb.split.us.split.us.split.us.split.us.split.us:
+; CHECK-NEXT: br label [[BB1_US_US_US_US_US:%.*]]
+; CHECK: bb1.us.us.us.us.us:
+; CHECK-NEXT: [[PHI_US_US_US_US_US:%.*]] = phi i64 [ 0, [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US]] ], [ [[ADD_US_US_US_US_US:%.*]], [[BB20_US_US_US_US_US:%.*]] ]
+; CHECK-NEXT: [[ADD_US_US_US_US_US]] = add nuw i64 [[PHI_US_US_US_US_US]], 1
+; CHECK-NEXT: [[ICMP_US_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP_US_US_US_US_US]], label [[BB2_US_US_US_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US:%.*]], !prof [[PROF0:![0-9]+]]
+; CHECK: bb2.us.us.us.us.us:
+; CHECK-NEXT: [[ICMP3_US_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB4_US_US_US_US_US:%.*]]
+; CHECK: bb4.us.us.us.us.us:
+; CHECK-NEXT: [[ICMP5_US_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB6_US_US_US_US_US:%.*]]
+; CHECK: bb6.us.us.us.us.us:
+; CHECK-NEXT: [[ICMP7_US_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB8_US_US_US_US_US:%.*]]
+; CHECK: bb8.us.us.us.us.us:
+; CHECK-NEXT: [[ICMP9_US_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB10_US_US_US_US_US:%.*]]
+; CHECK: bb10.us.us.us.us.us:
+; CHECK-NEXT: [[ICMP11_US_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB20_US_US_US_US_US]]
+; CHECK: bb20.us.us.us.us.us:
+; CHECK-NEXT: br label [[BB1_US_US_US_US_US]]
+; CHECK: bb21.split.us.split.us.split.us.split.us.split.us:
+; CHECK-NEXT: br label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US:%.*]]
+; CHECK: bb.split.us.split.us.split.us.split.us.split:
+; CHECK-NEXT: br label [[BB1_US_US_US_US:%.*]]
+; CHECK: bb1.us.us.us.us:
+; CHECK-NEXT: [[PHI_US_US_US_US:%.*]] = phi i64 [ 0, [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT]] ], [ [[ADD_US_US_US_US:%.*]], [[BB20_US_US_US_US:%.*]] ]
+; CHECK-NEXT: [[ADD_US_US_US_US]] = add nuw i64 [[PHI_US_US_US_US]], 1
+; CHECK-NEXT: [[ICMP_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP_US_US_US_US]], label [[BB2_US_US_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT:%.*]], !prof [[PROF0]]
+; CHECK: bb2.us.us.us.us:
+; CHECK-NEXT: [[ICMP3_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB4_US_US_US_US:%.*]]
+; CHECK: bb4.us.us.us.us:
+; CHECK-NEXT: [[ICMP5_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB6_US_US_US_US:%.*]]
+; CHECK: bb6.us.us.us.us:
+; CHECK-NEXT: [[ICMP7_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB8_US_US_US_US:%.*]]
+; CHECK: bb8.us.us.us.us:
+; CHECK-NEXT: [[ICMP9_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB10_US_US_US_US:%.*]]
+; CHECK: bb10.us.us.us.us:
+; CHECK-NEXT: [[ICMP11_US_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB10_US_US_US_US_CHECK:%.*]]
+; CHECK: bb10.us.us.us.us.check:
+; CHECK-NEXT: br i1 [[ICMP11_US_US_US_US]], label [[BB20_US_US_US_US]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT]]
+; CHECK: bb20.us.us.us.us:
+; CHECK-NEXT: br label [[BB1_US_US_US_US]], !llvm.loop [[LOOP1:![0-9]+]]
+; CHECK: bb21.split.us.split.us.split.us.split.us.split:
+; CHECK-NEXT: br label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT_US]]
+; CHECK: bb21.split.us.split.us.split.us.split.us:
+; CHECK-NEXT: br label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US:%.*]]
+; CHECK: bb.split.us.split.us.split.us.split:
+; CHECK-NEXT: br label [[BB1_US_US_US:%.*]]
+; CHECK: bb1.us.us.us:
+; CHECK-NEXT: [[PHI_US_US_US:%.*]] = phi i64 [ 0, [[BB_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT]] ], [ [[ADD_US_US_US:%.*]], [[BB20_US_US_US:%.*]] ]
+; CHECK-NEXT: [[ADD_US_US_US]] = add nuw i64 [[PHI_US_US_US]], 1
+; CHECK-NEXT: [[ICMP_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP_US_US_US]], label [[BB2_US_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT:%.*]], !prof [[PROF0]]
+; CHECK: bb2.us.us.us:
+; CHECK-NEXT: [[ICMP3_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB4_US_US_US:%.*]]
+; CHECK: bb4.us.us.us:
+; CHECK-NEXT: [[ICMP5_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB6_US_US_US:%.*]]
+; CHECK: bb6.us.us.us:
+; CHECK-NEXT: [[ICMP7_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB8_US_US_US:%.*]]
+; CHECK: bb8.us.us.us:
+; CHECK-NEXT: [[ICMP9_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB8_US_US_US_CHECK:%.*]]
+; CHECK: bb8.us.us.us.check:
+; CHECK-NEXT: br i1 [[ICMP9_US_US_US]], label [[BB10_US_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT]]
+; CHECK: bb10.us.us.us:
+; CHECK-NEXT: [[ICMP11_US_US_US:%.*]] = icmp ult i64 [[PHI_US_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP11_US_US_US]], label [[BB20_US_US_US]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US_SPLIT]], !prof [[PROF0]]
+; CHECK: bb20.us.us.us:
+; CHECK-NEXT: br label [[BB1_US_US_US]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: bb21.split.us.split.us.split.us.split:
+; CHECK-NEXT: br label [[BB21_SPLIT_US_SPLIT_US_SPLIT_US]]
+; CHECK: bb21.split.us.split.us.split.us:
+; CHECK-NEXT: br label [[BB21_SPLIT_US_SPLIT_US:%.*]]
+; CHECK: bb.split.us.split.us.split:
+; CHECK-NEXT: br label [[BB1_US_US:%.*]]
+; CHECK: bb1.us.us:
+; CHECK-NEXT: [[PHI_US_US:%.*]] = phi i64 [ 0, [[BB_SPLIT_US_SPLIT_US_SPLIT]] ], [ [[ADD_US_US:%.*]], [[BB20_US_US:%.*]] ]
+; CHECK-NEXT: [[ADD_US_US]] = add nuw i64 [[PHI_US_US]], 1
+; CHECK-NEXT: [[ICMP_US_US:%.*]] = icmp ult i64 [[PHI_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP_US_US]], label [[BB2_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT:%.*]], !prof [[PROF0]]
+; CHECK: bb2.us.us:
+; CHECK-NEXT: [[ICMP3_US_US:%.*]] = icmp ult i64 [[PHI_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB4_US_US:%.*]]
+; CHECK: bb4.us.us:
+; CHECK-NEXT: [[ICMP5_US_US:%.*]] = icmp ult i64 [[PHI_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB6_US_US:%.*]]
+; CHECK: bb6.us.us:
+; CHECK-NEXT: [[ICMP7_US_US:%.*]] = icmp ult i64 [[PHI_US_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB6_US_US_CHECK:%.*]]
+; CHECK: bb6.us.us.check:
+; CHECK-NEXT: br i1 [[ICMP7_US_US]], label [[BB8_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT]]
+; CHECK: bb8.us.us:
+; CHECK-NEXT: [[ICMP9_US_US:%.*]] = icmp ult i64 [[PHI_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP9_US_US]], label [[BB10_US_US:%.*]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT]], !prof [[PROF0]]
+; CHECK: bb10.us.us:
+; CHECK-NEXT: [[ICMP11_US_US:%.*]] = icmp ult i64 [[PHI_US_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP11_US_US]], label [[BB20_US_US]], label [[BB21_SPLIT_US_SPLIT_US_SPLIT]], !prof [[PROF0]]
+; CHECK: bb20.us.us:
+; CHECK-NEXT: br label [[BB1_US_US]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK: bb21.split.us.split.us.split:
+; CHECK-NEXT: br label [[BB21_SPLIT_US_SPLIT_US]]
+; CHECK: bb21.split.us.split.us:
+; CHECK-NEXT: br label [[BB21_SPLIT_US:%.*]]
+; CHECK: bb.split.us.split:
+; CHECK-NEXT: br label [[BB1_US:%.*]]
+; CHECK: bb1.us:
+; CHECK-NEXT: [[PHI_US:%.*]] = phi i64 [ 0, [[BB_SPLIT_US_SPLIT]] ], [ [[ADD_US:%.*]], [[BB20_US:%.*]] ]
+; CHECK-NEXT: [[ADD_US]] = add nuw i64 [[PHI_US]], 1
+; CHECK-NEXT: [[ICMP_US:%.*]] = icmp ult i64 [[PHI_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP_US]], label [[BB2_US:%.*]], label [[BB21_SPLIT_US_SPLIT:%.*]], !prof [[PROF0]]
+; CHECK: bb2.us:
+; CHECK-NEXT: [[ICMP3_US:%.*]] = icmp ult i64 [[PHI_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB4_US:%.*]]
+; CHECK: bb4.us:
+; CHECK-NEXT: [[ICMP5_US:%.*]] = icmp ult i64 [[PHI_US]], [[ARG]]
+; CHECK-NEXT: br label [[BB4_US_CHECK:%.*]]
+; CHECK: bb4.us.check:
+; CHECK-NEXT: br i1 [[ICMP5_US]], label [[BB6_US:%.*]], label [[BB22_SPLIT_US:%.*]]
+; CHECK: bb6.us:
+; CHECK-NEXT: [[ICMP7_US:%.*]] = icmp ult i64 [[PHI_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP7_US]], label [[BB8_US:%.*]], label [[BB21_SPLIT_US_SPLIT]], !prof [[PROF0]]
+; CHECK: bb8.us:
+; CHECK-NEXT: [[ICMP9_US:%.*]] = icmp ult i64 [[PHI_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP9_US]], label [[BB10_US:%.*]], label [[BB21_SPLIT_US_SPLIT]], !prof [[PROF0]]
+; CHECK: bb10.us:
+; CHECK-NEXT: [[ICMP11_US:%.*]] = icmp ult i64 [[PHI_US]], [[ARG]]
+; CHECK-NEXT: br i1 [[ICMP11_US]], label [[BB20_US]], label [[BB21_SPLIT_US_SPLIT]], !prof [[PROF0]]
+; CHECK: bb20.us:
+; CHECK-NEXT: br label [[BB1_US]], !llvm.loop [[LOOP5:![0-9]+]]
+; CHECK: bb21.split.us.split:
+; CHECK-NEXT: br label [[BB21_SPLIT_US]]
+; CHECK: bb21.split.us:
+; CHECK-NEXT: br label [[BB21:%.*]]
+; CHECK: bb22.split.us:
+; CHECK-NEXT: br label [[BB22:%.*]]
+; CHECK: bb.split:
+; CHECK-NEXT: br label [[BB1:%.*]]
+; CHECK: bb1:
+; CHECK-NEXT: [[PHI:%.*]] = phi i64 [ 0, [[BB_SPLIT]] ], [ [[ADD:%.*]], [[BB20:%.*]] ]
+; CHECK-NEXT: [[ADD]] = add nuw i64 [[PHI]], 1
+; CHECK-NEXT: [[ICMP:%.*]] = icmp u...
[truncated]
|
fhahn
approved these changes
Sep 20, 2023
| @@ -3768,7 +3775,7 @@ bool SimpleLoopUnswitchLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) { | |||
| // If the current loop has been unswitched using a partially invariant | |||
| // condition, we should not re-add the current loop to avoid unswitching | |||
| // on the same condition again. | |||
There was a problem hiding this comment.
Comment above needs updating
When unswitching via invariant condition injection, we currently mark the condition in the old loop, so that it does not get unswitched again. However, if there are multiple branches for which conditions can be injected, then we can do that for both the old and new loop. This means that the number of unswitches increases exponentially. Change the handling to be more similar to partial unswitching, where we instead mark the whole loop, rather than a single condition. This means that we will only generate a linear number of loops. TBH I think even that is still highly undesirable, and we should probably be unswitching all candidates at the same time, so that we end up with only two loops. But at least this mitigates the worst case. The test case is a reduced variant that generates 1700 lines of IR without this patch and 290 with it. Fixes llvm#66868.
nikic
force-pushed
the
simple-loop-unswitch-fix
branch
from
866577f to
8362cae
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When unswitching via invariant condition injection, we currently mark the condition in the old loop, so that it does not get unswitched again. However, if there are multiple branches for which conditions can be injected, then we can do that for both the old and new loops. This means that the number of unswitches increases exponentially.
Change the handling to be more similar to partial unswitching, where we instead mark the whole loop, rather than a single condition. This means that we will only generate a linear number of loops.
TBH I think even that is still highly undesirable, and we should probably be unswitching all candidates at the same time, so that we end up with only two loops. But at least this mitigates the worst case.
The test case is a reduced variant that generates 1700 lines of IR without this patch and 290 with it.
Fixes #66868.