GEP Loop Strength Reduction pass improvements

1. Use one instance of SCEVExpander for all reduction in the functions.
   Limits number of duplicated instructions.

2. Improve comparison of SCEV expressions. More cases should be grouped
   together, lowering number of duplicated instructions.

Author: pkwasnie-intel

IGC/Compiler/Optimizer/OpenCLPasses/GEPLoopStrengthReduction/GEPLoopStrengthReduction.cpp

@@ -262,8 +262,8 @@ class Scorer
 class Reducer
 {
 public:
-    Reducer(const DataLayout &DL, DominatorTree &DT, Loop &L, LoopInfo &LI, ModuleMetaData &MMD, RegisterPressureEstimate &RPE, ScalarEvolution &SE, unsigned MaxPressure, bool AllowLICM)
-        : DT(DT), L(L), LI(LI), RPE(RPE), SE(SE), E(SE, DL, "gep-loop-strength-reduction"), MaxPressure(MaxPressure), AllowLICM(AllowLICM), Scorer(DL, L, MMD) {}
+    Reducer(const DataLayout &DL, DominatorTree &DT, Loop &L, LoopInfo &LI, ModuleMetaData &MMD, RegisterPressureEstimate &RPE, ScalarEvolution &SE, SCEVExpander &E, unsigned MaxPressure, bool AllowLICM)
+        : DT(DT), L(L), LI(LI), RPE(RPE), SE(SE), E(E), MaxPressure(MaxPressure), AllowLICM(AllowLICM), Scorer(DL, L, MMD) {}
 
     bool reduce();
 
@@ -280,7 +280,7 @@ class Reducer
     LoopInfo &LI;
     RegisterPressureEstimate &RPE;
     ScalarEvolution &SE;
-    SCEVExpander E;
+    SCEVExpander &E;
 
     Scorer Scorer;
 
@@ -305,22 +305,26 @@ namespace SCEVHelper
     class SCEVAddBuilder
     {
     public:
-        SCEVAddBuilder(ScalarEvolution &SE, Type *T) :
-            SE(SE), T(T)
-        {
-            IGC_ASSERT_MESSAGE(T->isIntegerTy(), "builder requires integer type");
-        }
+        SCEVAddBuilder(ScalarEvolution &SE) : SE(SE) {}
 
         SCEVAddBuilder &add(const SCEV *S, bool Negative = false);
 
         SCEVAddBuilder &addNegative(const SCEV *S) { return add(S, true); }
 
-        const SCEV *build() { return SE.getAddExpr(Ops); }
+        const SCEV *build();
 
     private:
+
+        struct Op
+        {
+            Op(const SCEV *S, bool Negative) : S(S), Negative(Negative) {}
+
+            const SCEV *S;
+            bool Negative;
+        };
+
         ScalarEvolution &SE;
-        Type *T;
-        SmallVector<const SCEV*, 16> Ops;
+        SmallVector<Op, 16> Ops;
     };
 };
 
@@ -406,7 +410,7 @@ bool ReductionCandidateGroup::addToGroup(ScalarEvolution &SE, GetElementPtrInst
     // Can't use ScalarEvolution::computeConstantDifference, as it only
     // supports SCEVAddExpr with two operands. Calculate difference as:
     //     new candidate's operands + (-1 * base's operands)
-    SCEVHelper::SCEVAddBuilder Builder(SE, SE.getWiderType(S->getType(), Base.S->getType()));
+    SCEVHelper::SCEVAddBuilder Builder(SE);
     const SCEVConstant *Sum = dyn_cast<SCEVConstant>(Builder.add(S).addNegative(Base.S).build());
     if (!Sum)
         return false;
@@ -900,7 +904,7 @@ bool Reducer::deconstructSCEV(const SCEV *S, const SCEV *&Start, int64_t &Step)
 
         const SCEV *OpSCEV = nullptr;
         int64_t OpStep = 0;
-        SCEVHelper::SCEVAddBuilder Builder(SE, S->getType());
+        SCEVHelper::SCEVAddBuilder Builder(SE);
 
         for (auto *Op : Add->operands())
         {
@@ -1036,16 +1040,33 @@ SCEVHelper::SCEVAddBuilder &SCEVHelper::SCEVAddBuilder::add(const SCEV *S, bool
         return *this;
     }
 
+    Ops.emplace_back(S, Negative);
+
+    return *this;
+}
+
+
+const SCEV *SCEVHelper::SCEVAddBuilder::build()
+{
     // ScalarEvolution::getAddExpr requires all operands to have the same
-    // type. Extend type if required.
-    S = S->getType() == T ? S : SE.getSignExtendExpr(S, T);
+    // type. First find the widest type.
+    Type *T = nullptr;
+    for (auto *It = Ops.begin(); It != Ops.end(); ++It)
+    {
+        T = T ? SE.getWiderType(T, It->S->getType()) : It->S->getType();
+    }
 
-    // Change expresion to "-1 * expression"
-    S = Negative ? SE.getNegativeSCEV(S) : S;
+    // Join list of operands, extending type if required.
+    SmallVector<const SCEV*, 16> FinalOps;
 
-    Ops.push_back(S);
+    for (auto *It = Ops.begin(); It != Ops.end(); ++It)
+    {
+        const SCEV *S = It->S;
+        S = S->getType() == T ? S : SE.getSignExtendExpr(S, T);
+        FinalOps.push_back(It->Negative ? SE.getNegativeSCEV(S) : S);
+    }
 
-    return *this;
+    return SE.getAddExpr(FinalOps);
 }
 
 
@@ -1093,9 +1114,12 @@ bool GEPLoopStrengthReduction::runOnFunction(llvm::Function &F)
 
     bool changed = false;
 
+    // Using one SCEV expander between all reductions reduces number of duplicated new instructions.
+    auto E = SCEVExpander(SE, DL, "gep-loop-strength-reduction");
+
     for (Loop *L : LI.getLoopsInPreorder())
     {
-        changed |= Reducer(DL, DT, *L, LI, MMD, RPE, SE, MaxPressure, AllowLICM).reduce();
+        changed |= Reducer(DL, DT, *L, LI, MMD, RPE, SE, E, MaxPressure, AllowLICM).reduce();
     }
 
     return changed;

IGC/Compiler/tests/GEPLoopStrengthReduction/one_access_addexpr.ll

@@ -35,14 +35,10 @@ entry:
   br i1 %cmp1, label %for.body.lr.ph, label %for.end
 
 ; CHECK-LABEL: for.body.lr.ph:
-; CHECK:         [[EXT1:%.*]] = sext i32 %b to i64
-; CHECK:         [[EXT2:%.*]] = sext i32 %a to i64
-; CHECK:         [[ADD1:%.*]] = add i64 [[EXT1]], [[EXT2]]
-; CHECK:         [[ADD2:%.*]] = add i32 %b, %a
-; CHECK:         [[MUL:%.*]] = mul i32 %c, [[ADD2]]
-; CHECK:         [[EXT3:%.*]] = sext i32 [[MUL]] to i64
-; CHECK:         [[ADD3:%.*]] = add i64 [[ADD1]], [[EXT3]]
-; CHECK:         [[GEP_PHI1:%.*]] = getelementptr i32, i32 addrspace(1)* %p, i64 [[ADD3]]
+; CHECK:         [[ADD1:%.*]] = add i32 %b, %a
+; CHECK:         [[MUL:%.*]] = mul i32 %c, [[ADD1]]
+; CHECK:         [[ADD2:%.*]] = add i32 [[ADD1]], [[MUL]]
+; CHECK:         [[GEP_PHI1:%.*]] = getelementptr i32, i32 addrspace(1)* %p, i32 [[ADD2]]
 ; CHECK:         br label %for.body
 for.body.lr.ph:                                   ; preds = %entry
   br label %for.body