[SCEV] Improve handling of divisibility information from loop guards.

At the moment, the effectivness of guards that contain divisibility
information (A % B == 0 ) depends on the order of the conditions.

This patch makes using divisibility information independent of the
order, by collecting and applying the divisibility information
separately.

We first collect all conditions in a vector, then collect the
divisibility information from all guards.

When processing other guards, we apply divisibility info collected
earlier.

After all guards have been processed, we add the divisibility info,
rewriting the existing rewrite. This ensures we apply the divisibility
info to the largest rewrite expression.

This helps to improve results in a few cases, one in
dtcxzyw/llvm-opt-benchmark#2921 and another one
in a different large C/C++ based IR corpus.

Author: fhahn

llvm/lib/Analysis/ScalarEvolution.cpp

@@ -15474,11 +15474,26 @@ void ScalarEvolution::LoopGuards::collectFromPHI(
 }
 
 // Return a new SCEV that modifies \p Expr to the closest number divides by
-// \p Divisor and greater or equal than Expr. For now, only handle constant
+// \p Divisor and less or equal than Expr. For now, only handle constant
 // Expr.
-static const SCEV *getNextSCEVDividesByDivisor(const SCEV *Expr,
-                                               const APInt &DivisorVal,
-                                               ScalarEvolution &SE) {
+static const SCEV *getPreviousSCEVDivisibleByDivisor(const SCEV *Expr,
+                                                     const APInt &DivisorVal,
+                                                     ScalarEvolution &SE) {
+  const APInt *ExprVal;
+  if (!match(Expr, m_scev_APInt(ExprVal)) || ExprVal->isNegative() ||
+      DivisorVal.isNonPositive())
+    return Expr;
+  APInt Rem = ExprVal->urem(DivisorVal);
+  // return the SCEV: Expr - Expr % Divisor
+  return SE.getConstant(*ExprVal - Rem);
+}
+
+// Return a new SCEV that modifies \p Expr to the closest number divisible by
+// \p Divisor and greater than or equal to Expr.
+// For now, only handle constant Expr and Divisor.
+static const SCEV *getNextSCEVDivisibleByDivisor(const SCEV *Expr,
+                                                 const APInt &DivisorVal,
+                                                 ScalarEvolution &SE) {
   const APInt *ExprVal;
   if (!match(Expr, m_scev_APInt(ExprVal)) || ExprVal->isNegative() ||
       DivisorVal.isNonPositive())
@@ -15490,6 +15505,79 @@ static const SCEV *getNextSCEVDividesByDivisor(const SCEV *Expr,
   return SE.getConstant(*ExprVal + DivisorVal - Rem);
 }
 
+static bool collectDivisibilityInformation(
+    ICmpInst::Predicate Predicate, const SCEV *LHS, const SCEV *RHS,
+    DenseMap<const SCEV *, const SCEV *> &DivInfo,
+    DenseMap<const SCEV *, APInt> &Multiples, ScalarEvolution &SE) {
+  // If we have LHS == 0, check if LHS is computing a property of some unknown
+  // SCEV %v which we can rewrite %v to express explicitly.
+  if (Predicate != CmpInst::ICMP_EQ || !match(RHS, m_scev_Zero()))
+    return false;
+  // If LHS is A % B, i.e. A % B == 0, rewrite A to (A /u B) * B to
+  // explicitly express that.
+  const SCEVUnknown *URemLHS = nullptr;
+  const SCEV *URemRHS = nullptr;
+  if (!match(LHS, m_scev_URem(m_SCEVUnknown(URemLHS), m_SCEV(URemRHS), SE)))
+    return false;
+
+  const SCEV *Multiple =
+      SE.getMulExpr(SE.getUDivExpr(URemLHS, URemRHS), URemRHS);
+  DivInfo[URemLHS] = Multiple;
+  Multiples[URemLHS] = cast<SCEVConstant>(URemRHS)->getAPInt();
+  return true;
+}
+
+// Check if the condition is a divisibility guard (A % B == 0).
+static bool isDivisibilityGuard(const SCEV *LHS, const SCEV *RHS,
+                                ScalarEvolution &SE) {
+  const SCEV *X, *Y;
+  return match(LHS, m_scev_URem(m_SCEV(X), m_SCEV(Y), SE)) && RHS->isZero();
+}
+
+// Apply divisibility by \p Divisor on MinMaxExpr with constant values,
+// recursively. This is done by aligning up/down the constant value to the
+// Divisor.
+static const SCEV *applyDivisibilityOnMinMaxExpr(const SCEV *MinMaxExpr,
+                                                 const SCEV *Divisor,
+                                                 ScalarEvolution &SE) {
+  // Return true if \p Expr is a MinMax SCEV expression with a non-negative
+  // constant operand. If so, return in \p SCTy the SCEV type and in \p RHS
+  // the non-constant operand and in \p LHS the constant operand.
+  auto IsMinMaxSCEVWithNonNegativeConstant =
+      [&](const SCEV *Expr, SCEVTypes &SCTy, const SCEV *&LHS,
+          const SCEV *&RHS) {
+        if (auto *MinMax = dyn_cast<SCEVMinMaxExpr>(Expr)) {
+          if (MinMax->getNumOperands() != 2)
+            return false;
+          if (auto *C = dyn_cast<SCEVConstant>(MinMax->getOperand(0))) {
+            if (C->getAPInt().isNegative())
+              return false;
+            SCTy = MinMax->getSCEVType();
+            LHS = MinMax->getOperand(0);
+            RHS = MinMax->getOperand(1);
+            return true;
+          }
+        }
+        return false;
+      };
+
+  const SCEV *MinMaxLHS = nullptr, *MinMaxRHS = nullptr;
+  SCEVTypes SCTy;
+  if (!IsMinMaxSCEVWithNonNegativeConstant(MinMaxExpr, SCTy, MinMaxLHS,
+                                           MinMaxRHS))
+    return MinMaxExpr;
+  auto IsMin = isa<SCEVSMinExpr>(MinMaxExpr) || isa<SCEVUMinExpr>(MinMaxExpr);
+  assert(SE.isKnownNonNegative(MinMaxLHS) && "Expected non-negative operand!");
+  auto *DivisibleExpr =
+      IsMin ? getPreviousSCEVDivisibleByDivisor(
+                  MinMaxLHS, cast<SCEVConstant>(Divisor)->getAPInt(), SE)
+            : getNextSCEVDivisibleByDivisor(
+                  MinMaxLHS, cast<SCEVConstant>(Divisor)->getAPInt(), SE);
+  SmallVector<const SCEV *> Ops = {
+      applyDivisibilityOnMinMaxExpr(MinMaxRHS, Divisor, SE), DivisibleExpr};
+  return SE.getMinMaxExpr(SCTy, Ops);
+}
+
 void ScalarEvolution::LoopGuards::collectFromBlock(
     ScalarEvolution &SE, ScalarEvolution::LoopGuards &Guards,
     const BasicBlock *Block, const BasicBlock *Pred,
@@ -15500,19 +15588,14 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
   SmallVector<const SCEV *> ExprsToRewrite;
   auto CollectCondition = [&](ICmpInst::Predicate Predicate, const SCEV *LHS,
                               const SCEV *RHS,
-                              DenseMap<const SCEV *, const SCEV *>
-                                  &RewriteMap) {
+                              DenseMap<const SCEV *, const SCEV *> &RewriteMap,
+                              const DenseMap<const SCEV *, const SCEV *>
+                                  &DivInfo) {
     // WARNING: It is generally unsound to apply any wrap flags to the proposed
     // replacement SCEV which isn't directly implied by the structure of that
     // SCEV.  In particular, using contextual facts to imply flags is *NOT*
     // legal.  See the scoping rules for flags in the header to understand why.
 
-    // If LHS is a constant, apply information to the other expression.
-    if (isa<SCEVConstant>(LHS)) {
-      std::swap(LHS, RHS);
-      Predicate = CmpInst::getSwappedPredicate(Predicate);
-    }
-
     // Check for a condition of the form (-C1 + X < C2).  InstCombine will
     // create this form when combining two checks of the form (X u< C2 + C1) and
     // (X >=u C1).
@@ -15545,80 +15628,6 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
     if (MatchRangeCheckIdiom())
       return;
 
-    // Return true if \p Expr is a MinMax SCEV expression with a non-negative
-    // constant operand. If so, return in \p SCTy the SCEV type and in \p RHS
-    // the non-constant operand and in \p LHS the constant operand.
-    auto IsMinMaxSCEVWithNonNegativeConstant =
-        [&](const SCEV *Expr, SCEVTypes &SCTy, const SCEV *&LHS,
-            const SCEV *&RHS) {
-          const APInt *C;
-          SCTy = Expr->getSCEVType();
-          return match(Expr, m_scev_MinMax(m_SCEV(LHS), m_SCEV(RHS))) &&
-                 match(LHS, m_scev_APInt(C)) && C->isNonNegative();
-        };
-
-    // Return a new SCEV that modifies \p Expr to the closest number divides by
-    // \p Divisor and less or equal than Expr. For now, only handle constant
-    // Expr.
-    auto GetPreviousSCEVDividesByDivisor = [&](const SCEV *Expr,
-                                               const APInt &DivisorVal) {
-      const APInt *ExprVal;
-      if (!match(Expr, m_scev_APInt(ExprVal)) || ExprVal->isNegative() ||
-          DivisorVal.isNonPositive())
-        return Expr;
-      APInt Rem = ExprVal->urem(DivisorVal);
-      // return the SCEV: Expr - Expr % Divisor
-      return SE.getConstant(*ExprVal - Rem);
-    };
-
-    // Apply divisibilty by \p Divisor on MinMaxExpr with constant values,
-    // recursively. This is done by aligning up/down the constant value to the
-    // Divisor.
-    std::function<const SCEV *(const SCEV *, const SCEV *)>
-        ApplyDivisibiltyOnMinMaxExpr = [&](const SCEV *MinMaxExpr,
-                                           const SCEV *Divisor) {
-          auto *ConstDivisor = dyn_cast<SCEVConstant>(Divisor);
-          if (!ConstDivisor)
-            return MinMaxExpr;
-          const APInt &DivisorVal = ConstDivisor->getAPInt();
-
-          const SCEV *MinMaxLHS = nullptr, *MinMaxRHS = nullptr;
-          SCEVTypes SCTy;
-          if (!IsMinMaxSCEVWithNonNegativeConstant(MinMaxExpr, SCTy, MinMaxLHS,
-                                                   MinMaxRHS))
-            return MinMaxExpr;
-          auto IsMin =
-              isa<SCEVSMinExpr>(MinMaxExpr) || isa<SCEVUMinExpr>(MinMaxExpr);
-          assert(SE.isKnownNonNegative(MinMaxLHS) &&
-                 "Expected non-negative operand!");
-          auto *DivisibleExpr =
-              IsMin ? GetPreviousSCEVDividesByDivisor(MinMaxLHS, DivisorVal)
-                    : getNextSCEVDividesByDivisor(MinMaxLHS, DivisorVal, SE);
-          SmallVector<const SCEV *> Ops = {
-              ApplyDivisibiltyOnMinMaxExpr(MinMaxRHS, Divisor), DivisibleExpr};
-          return SE.getMinMaxExpr(SCTy, Ops);
-        };
-
-    // If we have LHS == 0, check if LHS is computing a property of some unknown
-    // SCEV %v which we can rewrite %v to express explicitly.
-    if (Predicate == CmpInst::ICMP_EQ && match(RHS, m_scev_Zero())) {
-      // If LHS is A % B, i.e. A % B == 0, rewrite A to (A /u B) * B to
-      // explicitly express that.
-      const SCEVUnknown *URemLHS = nullptr;
-      const SCEV *URemRHS = nullptr;
-      if (match(LHS,
-                m_scev_URem(m_SCEVUnknown(URemLHS), m_SCEV(URemRHS), SE))) {
-        auto I = RewriteMap.find(URemLHS);
-        const SCEV *RewrittenLHS = I != RewriteMap.end() ? I->second : URemLHS;
-        RewrittenLHS = ApplyDivisibiltyOnMinMaxExpr(RewrittenLHS, URemRHS);
-        const auto *Multiple =
-            SE.getMulExpr(SE.getUDivExpr(RewrittenLHS, URemRHS), URemRHS);
-        RewriteMap[URemLHS] = Multiple;
-        ExprsToRewrite.push_back(URemLHS);
-        return;
-      }
-    }
-
     // Do not apply information for constants or if RHS contains an AddRec.
     if (isa<SCEVConstant>(LHS) || SE.containsAddRecurrence(RHS))
       return;
@@ -15648,7 +15657,11 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
     };
 
     const SCEV *RewrittenLHS = GetMaybeRewritten(LHS);
-    const APInt &DividesBy = SE.getConstantMultiple(RewrittenLHS);
+    // Apply divisibility information when computing the constant multiple.
+    LoopGuards DivGuards(SE);
+    DivGuards.RewriteMap = DivInfo;
+    const APInt &DividesBy =
+        SE.getConstantMultiple(DivGuards.rewrite(RewrittenLHS));
 
     // Collect rewrites for LHS and its transitive operands based on the
     // condition.
@@ -15663,31 +15676,31 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
     // predicate.
     const SCEV *One = SE.getOne(RHS->getType());
     switch (Predicate) {
-      case CmpInst::ICMP_ULT:
-        if (RHS->getType()->isPointerTy())
-          return;
-        RHS = SE.getUMaxExpr(RHS, One);
-        [[fallthrough]];
-      case CmpInst::ICMP_SLT: {
-        RHS = SE.getMinusSCEV(RHS, One);
-        RHS = GetPreviousSCEVDividesByDivisor(RHS, DividesBy);
-        break;
-      }
-      case CmpInst::ICMP_UGT:
-      case CmpInst::ICMP_SGT:
-        RHS = SE.getAddExpr(RHS, One);
-        RHS = getNextSCEVDividesByDivisor(RHS, DividesBy, SE);
-        break;
-      case CmpInst::ICMP_ULE:
-      case CmpInst::ICMP_SLE:
-        RHS = GetPreviousSCEVDividesByDivisor(RHS, DividesBy);
-        break;
-      case CmpInst::ICMP_UGE:
-      case CmpInst::ICMP_SGE:
-        RHS = getNextSCEVDividesByDivisor(RHS, DividesBy, SE);
-        break;
-      default:
-        break;
+    case CmpInst::ICMP_ULT:
+      if (RHS->getType()->isPointerTy())
+        return;
+      RHS = SE.getUMaxExpr(RHS, One);
+      [[fallthrough]];
+    case CmpInst::ICMP_SLT: {
+      RHS = SE.getMinusSCEV(RHS, One);
+      RHS = getPreviousSCEVDivisibleByDivisor(RHS, DividesBy, SE);
+      break;
+    }
+    case CmpInst::ICMP_UGT:
+    case CmpInst::ICMP_SGT:
+      RHS = SE.getAddExpr(RHS, One);
+      RHS = getNextSCEVDivisibleByDivisor(RHS, DividesBy, SE);
+      break;
+    case CmpInst::ICMP_ULE:
+    case CmpInst::ICMP_SLE:
+      RHS = getPreviousSCEVDivisibleByDivisor(RHS, DividesBy, SE);
+      break;
+    case CmpInst::ICMP_UGE:
+    case CmpInst::ICMP_SGE:
+      RHS = getNextSCEVDivisibleByDivisor(RHS, DividesBy, SE);
+      break;
+    default:
+      break;
     }
 
     SmallVector<const SCEV *, 16> Worklist(1, LHS);
@@ -15738,7 +15751,7 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
       case CmpInst::ICMP_NE:
         if (match(RHS, m_scev_Zero())) {
           const SCEV *OneAlignedUp =
-              getNextSCEVDividesByDivisor(One, DividesBy, SE);
+              getNextSCEVDivisibleByDivisor(One, DividesBy, SE);
           To = SE.getUMaxExpr(FromRewritten, OneAlignedUp);
         } else {
           // LHS != RHS can be rewritten as (LHS - RHS) = UMax(1, LHS - RHS),
@@ -15833,8 +15846,11 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
 
   // Now apply the information from the collected conditions to
   // Guards.RewriteMap. Conditions are processed in reverse order, so the
-  // earliest conditions is processed first. This ensures the SCEVs with the
+  // earliest conditions is processed first, except guards with divisibility
+  // information, which are moved to the back. This ensures the SCEVs with the
   // shortest dependency chains are constructed first.
+  SmallVector<std::tuple<CmpInst::Predicate, const SCEV *, const SCEV *>>
+      GuardsToProcess;
   for (auto [Term, EnterIfTrue] : reverse(Terms)) {
     SmallVector<Value *, 8> Worklist;
     SmallPtrSet<Value *, 8> Visited;
@@ -15849,7 +15865,12 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
             EnterIfTrue ? Cmp->getPredicate() : Cmp->getInversePredicate();
         const auto *LHS = SE.getSCEV(Cmp->getOperand(0));
         const auto *RHS = SE.getSCEV(Cmp->getOperand(1));
-        CollectCondition(Predicate, LHS, RHS, Guards.RewriteMap);
+        // If LHS is a constant, apply information to the other expression.
+        if (isa<SCEVConstant>(LHS)) {
+          std::swap(LHS, RHS);
+          Predicate = CmpInst::getSwappedPredicate(Predicate);
+        }
+        GuardsToProcess.emplace_back(Predicate, LHS, RHS);
         continue;
       }
 
@@ -15862,6 +15883,30 @@ void ScalarEvolution::LoopGuards::collectFromBlock(
     }
   }
 
+  // Process divisibility guards in reverse order to populate DivInfo early.
+  DenseMap<const SCEV *, APInt> Multiples;
+  DenseMap<const SCEV *, const SCEV *> DivInfo;
+  for (const auto &[Predicate, LHS, RHS] : GuardsToProcess) {
+    if (!isDivisibilityGuard(LHS, RHS, SE))
+      continue;
+    collectDivisibilityInformation(Predicate, LHS, RHS, DivInfo, Multiples, SE);
+  }
+
+  for (const auto &[Predicate, LHS, RHS] : GuardsToProcess)
+    CollectCondition(Predicate, LHS, RHS, Guards.RewriteMap, DivInfo);
+
+  // Apply divisibility information last. This ensures it is applied to the
+  // outermost expression after other rewrites for the given value.
+  for (const auto &[K, V] : Multiples) {
+    const SCEV *DivisorSCEV = SE.getConstant(V);
+    Guards.RewriteMap[K] =
+        SE.getMulExpr(SE.getUDivExpr(applyDivisibilityOnMinMaxExpr(
+                                         Guards.rewrite(K), DivisorSCEV, SE),
+                                     DivisorSCEV),
+                      DivisorSCEV);
+    ExprsToRewrite.push_back(K);
+  }
+
   // Let the rewriter preserve NUW/NSW flags if the unsigned/signed ranges of
   // the replacement expressions are contained in the ranges of the replaced
   // expressions.
@@ -15965,7 +16010,7 @@ const SCEV *ScalarEvolution::LoopGuards::rewrite(const SCEV *Expr) const {
           if (LHS > RHS)
             std::swap(LHS, RHS);
           if (NotEqual.contains({LHS, RHS})) {
-            const SCEV *OneAlignedUp = getNextSCEVDividesByDivisor(
+            const SCEV *OneAlignedUp = getNextSCEVDivisibleByDivisor(
                 SE.getOne(S->getType()), SE.getConstantMultiple(S), SE);
             return SE.getUMaxExpr(OneAlignedUp, S);
           }

llvm/test/Transforms/IndVarSimplify/loop-guard-order.ll

@@ -114,7 +114,7 @@ define i32 @urem_order1(i32 %n) {
 ; CHECK:       [[LOOP]]:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[LOOP_PREHEADER]] ]
 ; CHECK-NEXT:    call void @foo()
-; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 3
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw i32 [[IV]], 3
 ; CHECK-NEXT:    [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[N]]
 ; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT_LOOPEXIT:.*]], label %[[LOOP]]
 ; CHECK:       [[EXIT_LOOPEXIT]]:
@@ -205,13 +205,12 @@ define i64 @test_loop_with_div_order_1(i64 %n) {
 ; CHECK-NEXT:    [[PARITY_CHECK:%.*]] = icmp eq i64 [[IS_ODD]], 0
 ; CHECK-NEXT:    br i1 [[PARITY_CHECK]], label %[[LOOP_PREHEADER:.*]], label %[[EXIT]]
 ; CHECK:       [[LOOP_PREHEADER]]:
-; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[UPPER_BOUND]], i64 1)
 ; CHECK-NEXT:    br label %[[LOOP:.*]]
 ; CHECK:       [[LOOP]]:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[LOOP_PREHEADER]] ]
 ; CHECK-NEXT:    [[DUMMY:%.*]] = load volatile i64, ptr null, align 8
 ; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
-; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp ne i64 [[IV_NEXT]], [[UMAX]]
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp ne i64 [[IV_NEXT]], [[UPPER_BOUND]]
 ; CHECK-NEXT:    br i1 [[EXITCOND]], label %[[LOOP]], label %[[EXIT_LOOPEXIT:.*]]
 ; CHECK:       [[EXIT_LOOPEXIT]]:
 ; CHECK-NEXT:    br label %[[EXIT]]