[LoopPeel] Remove known trip count restriction when peeling last. (#140792)

Remove the restriction that the loop must be known to execute at least 2
iterations when peeling the last iteration. If we cannot prove at least
2 iterations are executed, a check and branch to skip the peeled loop is
inserted.

PR: llvm/llvm-project#140792

Author: fhahn

llvm/include/llvm/Transforms/Utils/LoopPeel.h

@@ -44,7 +44,8 @@ gatherPeelingPreferences(Loop *L, ScalarEvolution &SE,
 void computePeelCount(Loop *L, unsigned LoopSize,
                       TargetTransformInfo::PeelingPreferences &PP,
                       unsigned TripCount, DominatorTree &DT,
-                      ScalarEvolution &SE, AssumptionCache *AC = nullptr,
+                      ScalarEvolution &SE, const TargetTransformInfo &TTI,
+                      AssumptionCache *AC = nullptr,
                       unsigned Threshold = UINT_MAX);
 
 } // end namespace llvm

llvm/lib/Transforms/Scalar/LoopUnrollPass.cpp

@@ -1014,7 +1014,7 @@ bool llvm::computeUnrollCount(
   }
 
   // 5th priority is loop peeling.
-  computePeelCount(L, LoopSize, PP, TripCount, DT, SE, AC, UP.Threshold);
+  computePeelCount(L, LoopSize, PP, TripCount, DT, SE, TTI, AC, UP.Threshold);
   if (PP.PeelCount) {
     UP.Runtime = false;
     UP.Count = 1;

llvm/lib/Transforms/Utils/LoopPeel.cpp

@@ -38,6 +38,7 @@
 #include "llvm/Transforms/Utils/Cloning.h"
 #include "llvm/Transforms/Utils/LoopSimplify.h"
 #include "llvm/Transforms/Utils/LoopUtils.h"
+#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
 #include "llvm/Transforms/Utils/ValueMapper.h"
 #include <algorithm>
 #include <cassert>
@@ -330,11 +331,7 @@ static unsigned peelToTurnInvariantLoadsDerefencebale(Loop &L,
 
 bool llvm::canPeelLastIteration(const Loop &L, ScalarEvolution &SE) {
   const SCEV *BTC = SE.getBackedgeTakenCount(&L);
-  // The loop must execute at least 2 iterations to guarantee that peeled
-  // iteration executes.
-  // TODO: Add checks during codegen.
-  if (isa<SCEVCouldNotCompute>(BTC) ||
-      !SE.isKnownPredicate(CmpInst::ICMP_UGT, BTC, SE.getZero(BTC->getType())))
+  if (isa<SCEVCouldNotCompute>(BTC))
     return false;
 
   // Check if the exit condition of the loop can be adjusted by the peeling
@@ -364,12 +361,18 @@ bool llvm::canPeelLastIteration(const Loop &L, ScalarEvolution &SE) {
 /// is known at the second-to-last.
 static bool shouldPeelLastIteration(Loop &L, CmpPredicate Pred,
                                     const SCEVAddRecExpr *LeftAR,
-                                    const SCEV *RightSCEV,
-                                    ScalarEvolution &SE) {
+                                    const SCEV *RightSCEV, ScalarEvolution &SE,
+                                    const TargetTransformInfo &TTI) {
   if (!canPeelLastIteration(L, SE))
     return false;
 
   const SCEV *BTC = SE.getBackedgeTakenCount(&L);
+  SCEVExpander Expander(SE, L.getHeader()->getDataLayout(), "loop-peel");
+  if (!SE.isKnownNonZero(BTC) &&
+      Expander.isHighCostExpansion(BTC, &L, SCEVCheapExpansionBudget, &TTI,
+                                   L.getLoopPredecessor()->getTerminator()))
+    return false;
+
   const SCEV *ValAtLastIter = LeftAR->evaluateAtIteration(BTC, SE);
   const SCEV *ValAtSecondToLastIter = LeftAR->evaluateAtIteration(
       SE.getMinusSCEV(BTC, SE.getOne(BTC->getType())), SE);
@@ -391,7 +394,8 @@ static bool shouldPeelLastIteration(Loop &L, CmpPredicate Pred,
 //      ..
 //   }
 static std::pair<unsigned, unsigned>
-countToEliminateCompares(Loop &L, unsigned MaxPeelCount, ScalarEvolution &SE) {
+countToEliminateCompares(Loop &L, unsigned MaxPeelCount, ScalarEvolution &SE,
+                         const TargetTransformInfo &TTI) {
   assert(L.isLoopSimplifyForm() && "Loop needs to be in loop simplify form");
   unsigned DesiredPeelCount = 0;
   unsigned DesiredPeelCountLast = 0;
@@ -479,7 +483,7 @@ countToEliminateCompares(Loop &L, unsigned MaxPeelCount, ScalarEvolution &SE) {
     const SCEV *Step = LeftAR->getStepRecurrence(SE);
     if (!PeelWhilePredicateIsKnown(NewPeelCount, IterVal, RightSCEV, Step,
                                    Pred)) {
-      if (shouldPeelLastIteration(L, Pred, LeftAR, RightSCEV, SE))
+      if (shouldPeelLastIteration(L, Pred, LeftAR, RightSCEV, SE, TTI))
         DesiredPeelCountLast = 1;
       return;
     }
@@ -593,8 +597,8 @@ static bool violatesLegacyMultiExitLoopCheck(Loop *L) {
 void llvm::computePeelCount(Loop *L, unsigned LoopSize,
                             TargetTransformInfo::PeelingPreferences &PP,
                             unsigned TripCount, DominatorTree &DT,
-                            ScalarEvolution &SE, AssumptionCache *AC,
-                            unsigned Threshold) {
+                            ScalarEvolution &SE, const TargetTransformInfo &TTI,
+                            AssumptionCache *AC, unsigned Threshold) {
   assert(LoopSize > 0 && "Zero loop size is not allowed!");
   // Save the PP.PeelCount value set by the target in
   // TTI.getPeelingPreferences or by the flag -unroll-peel-count.
@@ -656,7 +660,7 @@ void llvm::computePeelCount(Loop *L, unsigned LoopSize,
   }
 
   const auto &[CountToEliminateCmps, CountToEliminateCmpsLast] =
-      countToEliminateCompares(*L, MaxPeelCount, SE);
+      countToEliminateCompares(*L, MaxPeelCount, SE, TTI);
   DesiredPeelCount = std::max(DesiredPeelCount, CountToEliminateCmps);
 
   if (DesiredPeelCount == 0)
@@ -822,7 +826,7 @@ static void initBranchWeights(DenseMap<Instruction *, WeightInfo> &WeightInfos,
 /// instructions in the last peeled-off iteration.
 static void cloneLoopBlocks(
     Loop *L, unsigned IterNumber, bool PeelLast, BasicBlock *InsertTop,
-    BasicBlock *InsertBot,
+    BasicBlock *InsertBot, BasicBlock *OrigPreHeader,
     SmallVectorImpl<std::pair<BasicBlock *, BasicBlock *>> &ExitEdges,
     SmallVectorImpl<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks,
     ValueToValueMapTy &VMap, ValueToValueMapTy &LVMap, DominatorTree *DT,
@@ -914,12 +918,22 @@ static void cloneLoopBlocks(
   // loop iteration. Since this copy is no longer part of the loop, we
   // resolve this statically:
   if (PeelLast) {
-    // For the last iteration, we use the value from the latch of the original
-    // loop directly.
+    // For the last iteration, we introduce new phis for each header phi in
+    // InsertTop, using the incoming value from the preheader for the original
+    // preheader (when skipping the main loop) and the incoming value from the
+    // latch for the latch (when continuing from the main loop).
+    IRBuilder<> B(InsertTop->getTerminator());
     for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
       PHINode *NewPHI = cast<PHINode>(VMap[&*I]);
-      VMap[&*I] = NewPHI->getIncomingValueForBlock(Latch);
+      PHINode *PN = B.CreatePHI(NewPHI->getType(), 2);
       NewPHI->eraseFromParent();
+      if (OrigPreHeader)
+        PN->addIncoming(cast<PHINode>(&*I)->getIncomingValueForBlock(PreHeader),
+                        OrigPreHeader);
+
+      PN->addIncoming(cast<PHINode>(&*I)->getIncomingValueForBlock(Latch),
+                      Latch);
+      VMap[&*I] = PN;
     }
   } else {
     // For the first iteration, we use the value from the preheader directly.
@@ -1053,7 +1067,7 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
   // Set up all the necessary basic blocks.
   BasicBlock *InsertTop;
   BasicBlock *InsertBot;
-  BasicBlock *NewPreHeader;
+  BasicBlock *NewPreHeader = nullptr;
   DenseMap<Instruction *, Value *> ExitValues;
   if (PeelLast) {
     // It is convenient to split the single exit block from the latch the
@@ -1084,11 +1098,34 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
     for (PHINode &P : Exit->phis())
       ExitValues[&P] = P.getIncomingValueForBlock(Latch);
 
+    const SCEV *BTC = SE->getBackedgeTakenCount(L);
+
     InsertTop = SplitEdge(Latch, Exit, &DT, LI);
     InsertBot = SplitBlock(InsertTop, InsertTop->getTerminator(), &DT, LI);
 
     InsertTop->setName(Exit->getName() + ".peel.begin");
     InsertBot->setName(Exit->getName() + ".peel.next");
+    NewPreHeader = nullptr;
+
+    // If the original loop may only execute a single iteration we need to
+    // insert a trip count check and skip the original loop with the last
+    // iteration peeled off if necessary.
+    if (!SE->isKnownNonZero(BTC)) {
+      NewPreHeader = SplitEdge(PreHeader, Header, &DT, LI);
+      SCEVExpander Expander(*SE, Latch->getDataLayout(), "loop-peel");
+
+      BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
+      Value *BTCValue =
+          Expander.expandCodeFor(BTC, BTC->getType(), PreHeaderBR);
+      IRBuilder<> B(PreHeaderBR);
+      Value *Cond =
+          B.CreateICmpNE(BTCValue, ConstantInt::get(BTCValue->getType(), 0));
+      B.CreateCondBr(Cond, NewPreHeader, InsertTop);
+      PreHeaderBR->eraseFromParent();
+
+      // PreHeader now dominates InsertTop.
+      DT.changeImmediateDominator(InsertTop, PreHeader);
+    }
   } else {
     // It is convenient to split the preheader into 3 parts - two blocks to
     // anchor the peeled copy of the loop body, and a new preheader for the
@@ -1162,8 +1199,9 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
   for (unsigned Iter = 0; Iter < PeelCount; ++Iter) {
     SmallVector<BasicBlock *, 8> NewBlocks;
 
-    cloneLoopBlocks(L, Iter, PeelLast, InsertTop, InsertBot, ExitEdges,
-                    NewBlocks, LoopBlocks, VMap, LVMap, &DT, LI,
+    cloneLoopBlocks(L, Iter, PeelLast, InsertTop, InsertBot,
+                    NewPreHeader ? PreHeader : nullptr, ExitEdges, NewBlocks,
+                    LoopBlocks, VMap, LVMap, &DT, LI,
                     LoopLocalNoAliasDeclScopes, *SE);
 
     // Remap to use values from the current iteration instead of the
@@ -1216,9 +1254,11 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
 
   if (PeelLast) {
     // Now adjust users of the original exit values by replacing them with the
-    // exit value from the peeled iteration.
-    for (const auto &[P, E] : ExitValues)
+    // exit value from the peeled iteration and remove them.
+    for (const auto &[P, E] : ExitValues) {
       P->replaceAllUsesWith(isa<Constant>(E) ? E : &*VMap.lookup(E));
+      P->eraseFromParent();
+    }
     formLCSSA(*L, DT, LI, SE);
   } else {
     // Now adjust the phi nodes in the loop header to get their initial values

llvm/test/Transforms/LoopUnroll/peel-last-iteration-expansion-cost.ll

@@ -25,17 +25,41 @@ define i32 @test_expansion_cost_2(i32 %start, i32 %end) {
 ; BUDGET3-SAME: i32 [[START:%.*]], i32 [[END:%.*]]) {
 ; BUDGET3-NEXT:  [[ENTRY:.*]]:
 ; BUDGET3-NEXT:    [[SUB:%.*]] = add i32 [[END]], -1
+; BUDGET3-NEXT:    [[TMP0:%.*]] = sub i32 [[SUB]], [[START]]
+; BUDGET3-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
+; BUDGET3-NEXT:    br i1 [[TMP1]], label %[[ENTRY_SPLIT:.*]], label %[[EXIT_PEEL_BEGIN:.*]]
+; BUDGET3:       [[ENTRY_SPLIT]]:
 ; BUDGET3-NEXT:    br label %[[LOOP_HEADER:.*]]
 ; BUDGET3:       [[LOOP_HEADER]]:
-; BUDGET3-NEXT:    [[TMP3:%.*]] = phi i32 [ [[START]], %[[ENTRY]] ], [ [[IV_NEXT_PEEL:%.*]], %[[LOOP_LATCH:.*]] ]
-; BUDGET3-NEXT:    [[C_PEEL:%.*]] = icmp eq i32 [[TMP3]], [[SUB]]
-; BUDGET3-NEXT:    br i1 [[C_PEEL]], label %[[THEN:.*]], label %[[LOOP_LATCH]]
+; BUDGET3-NEXT:    [[IV:%.*]] = phi i32 [ [[START]], %[[ENTRY_SPLIT]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
+; BUDGET3-NEXT:    [[C:%.*]] = icmp eq i32 [[IV]], [[SUB]]
+; BUDGET3-NEXT:    br i1 [[C]], label %[[THEN:.*]], label %[[LOOP_LATCH]]
 ; BUDGET3:       [[THEN]]:
 ; BUDGET3-NEXT:    br label %[[LOOP_LATCH]]
 ; BUDGET3:       [[LOOP_LATCH]]:
-; BUDGET3-NEXT:    [[IV_NEXT_PEEL]] = add nsw i32 [[TMP3]], 1
+; BUDGET3-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
+; BUDGET3-NEXT:    [[TMP2:%.*]] = sub i32 [[END]], 1
+; BUDGET3-NEXT:    [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[TMP2]]
+; BUDGET3-NEXT:    br i1 [[EC]], label %[[EXIT_PEEL_BEGIN_LOOPEXIT:.*]], label %[[LOOP_HEADER]], !llvm.loop [[LOOP0:![0-9]+]]
+; BUDGET3:       [[EXIT_PEEL_BEGIN_LOOPEXIT]]:
+; BUDGET3-NEXT:    [[DOTPH:%.*]] = phi i32 [ [[IV_NEXT]], %[[LOOP_LATCH]] ]
+; BUDGET3-NEXT:    br label %[[EXIT_PEEL_BEGIN]]
+; BUDGET3:       [[EXIT_PEEL_BEGIN]]:
+; BUDGET3-NEXT:    [[TMP3:%.*]] = phi i32 [ [[START]], %[[ENTRY]] ], [ [[DOTPH]], %[[EXIT_PEEL_BEGIN_LOOPEXIT]] ]
+; BUDGET3-NEXT:    br label %[[LOOP_HEADER_PEEL:.*]]
+; BUDGET3:       [[LOOP_HEADER_PEEL]]:
+; BUDGET3-NEXT:    [[C_PEEL:%.*]] = icmp eq i32 [[TMP3]], [[SUB]]
+; BUDGET3-NEXT:    br i1 [[C_PEEL]], label %[[THEN_PEEL:.*]], label %[[LOOP_LATCH_PEEL:.*]]
+; BUDGET3:       [[THEN_PEEL]]:
+; BUDGET3-NEXT:    br label %[[LOOP_LATCH_PEEL]]
+; BUDGET3:       [[LOOP_LATCH_PEEL]]:
+; BUDGET3-NEXT:    [[IV_NEXT_PEEL:%.*]] = add nsw i32 [[TMP3]], 1
 ; BUDGET3-NEXT:    [[EC_PEEL:%.*]] = icmp eq i32 [[IV_NEXT_PEEL]], [[END]]
-; BUDGET3-NEXT:    br i1 [[EC_PEEL]], label %[[EXIT:.*]], label %[[LOOP_HEADER]]
+; BUDGET3-NEXT:    br i1 [[EC_PEEL]], label %[[EXIT_PEEL_NEXT:.*]], label %[[EXIT_PEEL_NEXT]]
+; BUDGET3:       [[EXIT_PEEL_NEXT]]:
+; BUDGET3-NEXT:    br label %[[LOOP_HEADER_PEEL_NEXT:.*]]
+; BUDGET3:       [[LOOP_HEADER_PEEL_NEXT]]:
+; BUDGET3-NEXT:    br label %[[EXIT:.*]]
 ; BUDGET3:       [[EXIT]]:
 ; BUDGET3-NEXT:    ret i32 0
 ;
@@ -59,3 +83,7 @@ loop.latch:
 exit:
   ret i32 0
 }
+;.
+; BUDGET3: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]]}
+; BUDGET3: [[META1]] = !{!"llvm.loop.peeled.count", i32 1}
+;.

llvm/test/Transforms/LoopUnroll/peel-last-iteration-with-constant-trip-count.ll

@@ -13,7 +13,6 @@ define i64 @peel_single_block_loop_iv_step_1() {
 ; CHECK-NEXT:    br i1 [[EC1]], label %[[LOOP]], label %[[EXIT_PEEL_BEGIN:.*]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[EXIT_PEEL_BEGIN]]:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT1]], %[[LOOP]] ]
-; CHECK-NEXT:    [[IV_LCSSA:%.*]] = phi i64 [ [[IV1]], %[[LOOP]] ]
 ; CHECK-NEXT:    br label %[[LOOP_PEEL:.*]]
 ; CHECK:       [[LOOP_PEEL]]:
 ; CHECK-NEXT:    [[CMP18_NOT:%.*]] = icmp eq i64 [[IV]], 63
@@ -92,7 +91,6 @@ define i64 @peel_single_block_loop_iv_step_1_eq_pred() {
 ; CHECK-NEXT:    br i1 [[CMP_PEEL]], label %[[EXIT_PEEL_BEGIN:.*]], label %[[LOOP]], !llvm.loop [[LOOP2:![0-9]+]]
 ; CHECK:       [[EXIT_PEEL_BEGIN]]:
 ; CHECK-NEXT:    [[IV_NEXT_LCSSA:%.*]] = phi i64 [ [[IV_LCSSA]], %[[LOOP]] ]
-; CHECK-NEXT:    [[IV_LCSSA1:%.*]] = phi i64 [ [[IV]], %[[LOOP]] ]
 ; CHECK-NEXT:    br label %[[LOOP_PEEL:.*]]
 ; CHECK:       [[LOOP_PEEL]]:
 ; CHECK-NEXT:    [[CMP_PEEL1:%.*]] = icmp eq i64 [[IV_NEXT_LCSSA]], 63
@@ -171,7 +169,6 @@ define i64 @peel_single_block_loop_iv_step_1_nested_loop() {
 ; CHECK-NEXT:    br i1 [[EC]], label %[[LOOP]], label %[[OUTER_LATCH_PEEL_BEGIN:.*]], !llvm.loop [[LOOP3:![0-9]+]]
 ; CHECK:       [[OUTER_LATCH_PEEL_BEGIN]]:
 ; CHECK-NEXT:    [[IV_NEXT_LCSSA:%.*]] = phi i64 [ [[IV_NEXT]], %[[LOOP]] ]
-; CHECK-NEXT:    [[IV_LCSSA:%.*]] = phi i64 [ [[IV]], %[[LOOP]] ]
 ; CHECK-NEXT:    br label %[[LOOP_PEEL:.*]]
 ; CHECK:       [[LOOP_PEEL]]:
 ; CHECK-NEXT:    [[CMP_PEEL:%.*]] = icmp eq i64 [[IV_NEXT_LCSSA]], 63
@@ -237,7 +234,6 @@ define i64 @peel_multi_block_loop_iv_step_1() {
 ; CHECK-NEXT:    br i1 [[EC]], label %[[LOOP]], label %[[EXIT_PEEL_BEGIN:.*]], !llvm.loop [[LOOP4:![0-9]+]]
 ; CHECK:       [[EXIT_PEEL_BEGIN]]:
 ; CHECK-NEXT:    [[IV_NEXT_LCSSA:%.*]] = phi i64 [ [[IV_NEXT]], %[[LATCH]] ]
-; CHECK-NEXT:    [[IV_LCSSA:%.*]] = phi i64 [ [[IV]], %[[LATCH]] ]
 ; CHECK-NEXT:    br label %[[LOOP_PEEL:.*]]
 ; CHECK:       [[LOOP_PEEL]]:
 ; CHECK-NEXT:    [[CMP_PEEL:%.*]] = icmp eq i64 [[IV_NEXT_LCSSA]], 63
@@ -365,7 +361,6 @@ define i64 @peel_single_block_loop_iv_step_1_btc_1() {
 ; CHECK-NEXT:    br i1 false, label %[[LOOP]], label %[[EXIT_PEEL_BEGIN:.*]], !llvm.loop [[LOOP5:![0-9]+]]
 ; CHECK:       [[EXIT_PEEL_BEGIN]]:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT1]], %[[LOOP]] ]
-; CHECK-NEXT:    [[IV_LCSSA:%.*]] = phi i64 [ [[IV1]], %[[LOOP]] ]
 ; CHECK-NEXT:    br label %[[LOOP_PEEL:.*]]
 ; CHECK:       [[LOOP_PEEL]]:
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[IV]], 1
@@ -483,9 +478,8 @@ define i32 @peel_loop_with_branch_and_phi_uses(ptr %x, i1 %c) {
 ; CHECK-NEXT:    [[EC1:%.*]] = icmp ne i32 [[IV_NEXT1]], 99
 ; CHECK-NEXT:    br i1 [[EC1]], label %[[LOOP_HEADER]], label %[[LOOPEXIT_PEEL_BEGIN:.*]], !llvm.loop [[LOOP6:![0-9]+]]
 ; CHECK:       [[LOOPEXIT_PEEL_BEGIN]]:
-; CHECK-NEXT:    [[RED:%.*]] = phi i32 [ [[ADD1]], %[[LOOP_LATCH]] ]
 ; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[IV_NEXT1]], %[[LOOP_LATCH]] ]
-; CHECK-NEXT:    [[ADD_LCSSA:%.*]] = phi i32 [ [[ADD1]], %[[LOOP_LATCH]] ]
+; CHECK-NEXT:    [[RED:%.*]] = phi i32 [ [[ADD1]], %[[LOOP_LATCH]] ]
 ; CHECK-NEXT:    br label %[[LOOP_HEADER_PEEL:.*]]
 ; CHECK:       [[LOOP_HEADER_PEEL]]:
 ; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[IV]], 99