[FuncSpec] Update function specialization to handle phi-chains

llvm/include/llvm/Transforms/IPO/FunctionSpecialization.h

@@ -217,6 +217,9 @@ class InstCostVisitor : public InstVisitor<InstCostVisitor, Constant *> {
   Cost estimateSwitchInst(SwitchInst &I);
   Cost estimateBranchInst(BranchInst &I);
 
+  void discoverTransitivelyIncomngValues(DenseSet<PHINode *> &PhiNodes,
+                                         PHINode *PN, unsigned Depth);
+
   Constant *visitInstruction(Instruction &I) { return nullptr; }
   Constant *visitPHINode(PHINode &I);
   Constant *visitFreezeInst(FreezeInst &I);

llvm/lib/Transforms/IPO/FunctionSpecialization.cpp

@@ -39,10 +39,15 @@ static cl::opt<unsigned> MaxClones(
     "The maximum number of clones allowed for a single function "
     "specialization"));
 
+static cl::opt<unsigned> MaxDiscoveryDepth(
+    "funcspec-max-discovery-depth", cl::init(10), cl::Hidden,
+    cl::desc("The maximum recursion depth allowed when searching for strongly "
+             "connected phis"));
+
 static cl::opt<unsigned> MaxIncomingPhiValues(
-    "funcspec-max-incoming-phi-values", cl::init(4), cl::Hidden, cl::desc(
-    "The maximum number of incoming values a PHI node can have to be "
-    "considered during the specialization bonus estimation"));
+    "funcspec-max-incoming-phi-values", cl::init(8), cl::Hidden,
+    cl::desc("The maximum number of incoming values a PHI node can have to be "
+             "considered during the specialization bonus estimation"));
 
 static cl::opt<unsigned> MaxBlockPredecessors(
     "funcspec-max-block-predecessors", cl::init(2), cl::Hidden, cl::desc(
@@ -64,9 +69,9 @@ static cl::opt<unsigned> MinCodeSizeSavings(
     "much percent of the original function size"));
 
 static cl::opt<unsigned> MinLatencySavings(
-    "funcspec-min-latency-savings", cl::init(70), cl::Hidden, cl::desc(
-    "Reject specializations whose latency savings are less than this"
-    "much percent of the original function size"));
+    "funcspec-min-latency-savings", cl::init(45), cl::Hidden,
+    cl::desc("Reject specializations whose latency savings are less than this"
+             "much percent of the original function size"));
 
 static cl::opt<unsigned> MinInliningBonus(
     "funcspec-min-inlining-bonus", cl::init(300), cl::Hidden, cl::desc(
@@ -262,30 +267,163 @@ Cost InstCostVisitor::estimateBranchInst(BranchInst &I) {
   return estimateBasicBlocks(WorkList);
 }
 
+// This function is finding candidates for a PHINode is part of a chain or graph
+// of PHINodes that all link to each other. That means, if the original input to
+// the chain is a constant all the other values are also that constant.
+//
+// The caller of this function will later check that no other nodes are involved
+// that are non-constant, and discard it from the possible conversions.
+//
+// For example:
+//
+// %a = load %0
+// %c = phi [%a, %d]
+// %d = phi [%e, %c]
+// %e = phi [%c, %f]
+// %f = phi [%j, %h]
+// %j = phi [%h, %j]
+// %h = phi [%g, %c]
+//
+// This is only showing the PHINodes, not the branches that choose the
+// different paths.
+//
+// A depth limit is used to avoid extreme recurusion.
+// A max number of incoming phi values ensures that expensive searches
+// are avoided.
+void InstCostVisitor::discoverTransitivelyIncomngValues(
+    DenseSet<PHINode *> &PHINodes, PHINode *PN, unsigned Depth) {
+  if (Depth > MaxDiscoveryDepth) {
+    LLVM_DEBUG(dbgs() << "FnSpecialization: Discover PHI nodes too deep ("
+                      << Depth << ">" << MaxDiscoveryDepth << ")\n");
+    return;
+  }
+
+  if (PN->getNumIncomingValues() > MaxIncomingPhiValues) {
+    LLVM_DEBUG(
+        dbgs() << "FnSpecialization: Discover PHI nodes has too many values  ("
+               << PN->getNumIncomingValues() << ">" << MaxIncomingPhiValues
+               << ")\n");
+    return;
+  }
+
+  // Already seen this, no more processing needed.
+  if (!PHINodes.insert(PN).second)
+    return;
+
+  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
+    Value *V = PN->getIncomingValue(I);
+    if (auto *Phi = dyn_cast<PHINode>(V)) {
+      if (Phi == PN || DeadBlocks.contains(PN->getIncomingBlock(I)))
+        continue;
+      discoverTransitivelyIncomngValues(PHINodes, Phi, Depth + 1);
+    }
+  }
+}
+
 Constant *InstCostVisitor::visitPHINode(PHINode &I) {
   if (I.getNumIncomingValues() > MaxIncomingPhiValues)
     return nullptr;
 
+  // PHI nodes
+  DenseSet<PHINode *> TransitivePHIs;
+
   bool Inserted = VisitedPHIs.insert(&I).second;
-  Constant *Const = nullptr;
+  SmallVector<PHINode *, 8> UnknownIncomingValues;
+
+  auto canConstantFoldPhiTrivially = [&](PHINode *PN) -> Constant * {
+    Constant *Const = nullptr;
 
-  for (unsigned Idx = 0, E = I.getNumIncomingValues(); Idx != E; ++Idx) {
-    Value *V = I.getIncomingValue(Idx);
-    if (auto *Inst = dyn_cast<Instruction>(V))
-      if (Inst == &I || DeadBlocks.contains(I.getIncomingBlock(Idx)))
+    UnknownIncomingValues.clear();
+    for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
+      Value *V = PN->getIncomingValue(I);
+
+      // Disregard self-references and dead incoming values.
+      if (auto *Inst = dyn_cast<Instruction>(V))
+        if (Inst == PN || DeadBlocks.contains(PN->getIncomingBlock(I)))
+          continue;
+
+      if (Constant *C = findConstantFor(V, KnownConstants)) {
+        if (!Const)
+          Const = C;
+        // Not all incoming values are the same constant. Bail immediately.
+        if (C != Const)
+          return nullptr;
         continue;
-    Constant *C = findConstantFor(V, KnownConstants);
-    if (!C) {
-      if (Inserted)
-        PendingPHIs.push_back(&I);
+      }
+      if (auto *Phi = dyn_cast<PHINode>(V)) {
+        UnknownIncomingValues.push_back(Phi);
+        continue;
+      }
+
+      // We can't reason about anything else.
       return nullptr;
     }
-    if (!Const)
-      Const = C;
-    else if (C != Const)
+    return UnknownIncomingValues.empty() ? Const : nullptr;
+  };
+
+  if (Constant *Const = canConstantFoldPhiTrivially(&I))
+    return Const;
+
+  if (Inserted) {
+    // First time we are seeing this phi. We'll retry later, after all
+    // the constant arguments have been propagated. Bail for now.
+    PendingPHIs.push_back(&I);
+    return nullptr;
+  }
+
+  // Try to see if we can collect a nest of transitive phis. Bail if
+  // it's too complex.
+  for (PHINode *Phi : UnknownIncomingValues)
+    discoverTransitivelyIncomngValues(TransitivePHIs, Phi, 1);
+
+
+  // A nested set of PHINodes can be constantfolded if:
+  // - It has a constant input.
+  // - It is always the SAME constant.
+  auto canConstantFoldNestedPhi = [&](PHINode *PN) -> Constant * {
+    Constant *Const = nullptr;
+
+    for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
+      Value *V = PN->getIncomingValue(I);
+
+      // Disregard self-references and dead incoming values.
+      if (auto *Inst = dyn_cast<Instruction>(V))
+        if (Inst == PN || DeadBlocks.contains(PN->getIncomingBlock(I)))
+          continue;
+
+      if (Constant *C = findConstantFor(V, KnownConstants)) {
+        if (!Const)
+          Const = C;
+        // Not all incoming values are the same constant. Bail immediately.
+        if (C != Const)
+          return nullptr;
+        continue;
+      }
+      if (auto *Phi = dyn_cast<PHINode>(V)) {
+        // It's not a Transitive phi. Bail out.
+        if (!TransitivePHIs.contains(Phi))
+          return nullptr;
+        continue;
+      }
+
+      // We can't reason about anything else.
       return nullptr;
+    }
+    return Const;
+  };
+
+  // All TransitivePHIs have to be the SAME constant.
+  Constant *Retval = nullptr;
+  for (PHINode *Phi : TransitivePHIs) {
+    if (Constant *Const = canConstantFoldNestedPhi(Phi)) {
+      if (!Retval)
+        Retval = Const;
+      else if (Retval != Const)
+        return nullptr;
+    }
   }
-  return Const;
+
+  return Retval;
 }
 
 Constant *InstCostVisitor::visitFreezeInst(FreezeInst &I) {
@@ -809,20 +947,41 @@ bool FunctionSpecializer::findSpecializations(Function *F, unsigned FuncSize,
       auto IsProfitable = [](Bonus &B, unsigned Score, unsigned FuncSize,
                              unsigned FuncGrowth) -> bool {
         // No check required.
-        if (ForceSpecialization)
+        if (ForceSpecialization) {
+          LLVM_DEBUG(dbgs() << "FnSpecialization: Force is on\n");
           return true;
+        }
         // Minimum inlining bonus.
-        if (Score > MinInliningBonus * FuncSize / 100)
+        if (Score > MinInliningBonus * FuncSize / 100) {
+          LLVM_DEBUG(dbgs()
+                     << "FnSpecialization: Sufficient inlining bonus (" << Score
+                     << " > " << MinInliningBonus * FuncSize / 100 << ")\n");
           return true;
+        }
         // Minimum codesize savings.
-        if (B.CodeSize < MinCodeSizeSavings * FuncSize / 100)
+        if (B.CodeSize < MinCodeSizeSavings * FuncSize / 100) {
+          LLVM_DEBUG(dbgs()
+                     << "FnSpecialization: Insufficient CodeSize Savings ("
+                     << B.CodeSize << " < "
+                     << MinCodeSizeSavings * FuncSize / 100 << ")\n");
           return false;
+        }
         // Minimum latency savings.
-        if (B.Latency < MinLatencySavings * FuncSize / 100)
+        if (B.Latency < MinLatencySavings * FuncSize / 100) {
+          LLVM_DEBUG(dbgs()
+                     << "FnSpecialization: Insufficient Latency Savings ("
+                     << B.Latency << " < " << MinLatencySavings * FuncSize / 100
+                     << ")\n");
           return false;
+        }
         // Maximum codesize growth.
-        if (FuncGrowth / FuncSize > MaxCodeSizeGrowth)
+        if (FuncGrowth / FuncSize > MaxCodeSizeGrowth) {
+          LLVM_DEBUG(dbgs()
+                     << "FnSpecialization: Function Growth exceeds threshold ("
+                     << FuncGrowth / FuncSize << " > " << MaxCodeSizeGrowth
+                     << ")\n");
           return false;
+        }
         return true;
       };
 

llvm/test/Transforms/FunctionSpecialization/discover-strongly-connected-phis.ll

@@ -0,0 +1,87 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+;
+; RUN: opt -passes="ipsccp<func-spec>" -funcspec-min-function-size=20 -funcspec-for-literal-constant -S < %s | FileCheck %s --check-prefix=FUNCSPEC
+; RUN: opt -passes="ipsccp<func-spec>" -funcspec-min-function-size=20 -funcspec-for-literal-constant -funcspec-max-discovery-depth=5 -S < %s | FileCheck %s --check-prefix=NOFUNCSPEC
+
+define i64 @bar(i1 %c1, i1 %c2, i1 %c3, i1 %c4, i1 %c5, i1 %c6, i1 %c7, i1 %c8, i1 %c9, i1 %c10) {
+; FUNCSPEC-LABEL: define i64 @bar(
+; FUNCSPEC-SAME: i1 [[C1:%.*]], i1 [[C2:%.*]], i1 [[C3:%.*]], i1 [[C4:%.*]], i1 [[C5:%.*]], i1 [[C6:%.*]], i1 [[C7:%.*]], i1 [[C8:%.*]], i1 [[C9:%.*]], i1 [[C10:%.*]]) {
+; FUNCSPEC-NEXT:  entry:
+; FUNCSPEC-NEXT:    [[F1:%.*]] = call i64 @foo.specialized.1(i64 3, i1 [[C1]], i1 [[C2]], i1 [[C3]], i1 [[C4]], i1 [[C5]], i1 [[C6]], i1 [[C7]], i1 [[C8]], i1 [[C9]], i1 [[C10]]), !range [[RNG0:![0-9]+]]
+; FUNCSPEC-NEXT:    [[F2:%.*]] = call i64 @foo.specialized.2(i64 4, i1 [[C1]], i1 [[C2]], i1 [[C3]], i1 [[C4]], i1 [[C5]], i1 [[C6]], i1 [[C7]], i1 [[C8]], i1 [[C9]], i1 [[C10]]), !range [[RNG1:![0-9]+]]
+; FUNCSPEC-NEXT:    [[ADD:%.*]] = add nuw nsw i64 [[F1]], [[F2]]
+; FUNCSPEC-NEXT:    ret i64 [[ADD]]
+;
+; NOFUNCSPEC-LABEL: define i64 @bar(
+; NOFUNCSPEC-SAME: i1 [[C1:%.*]], i1 [[C2:%.*]], i1 [[C3:%.*]], i1 [[C4:%.*]], i1 [[C5:%.*]], i1 [[C6:%.*]], i1 [[C7:%.*]], i1 [[C8:%.*]], i1 [[C9:%.*]], i1 [[C10:%.*]]) {
+; NOFUNCSPEC-NEXT:  entry:
+; NOFUNCSPEC-NEXT:    [[F1:%.*]] = call i64 @foo(i64 3, i1 [[C1]], i1 [[C2]], i1 [[C3]], i1 [[C4]], i1 [[C5]], i1 [[C6]], i1 [[C7]], i1 [[C8]], i1 [[C9]], i1 [[C10]]), !range [[RNG0:![0-9]+]]
+; NOFUNCSPEC-NEXT:    [[F2:%.*]] = call i64 @foo(i64 4, i1 [[C1]], i1 [[C2]], i1 [[C3]], i1 [[C4]], i1 [[C5]], i1 [[C6]], i1 [[C7]], i1 [[C8]], i1 [[C9]], i1 [[C10]]), !range [[RNG0]]
+; NOFUNCSPEC-NEXT:    [[ADD:%.*]] = add nuw nsw i64 [[F1]], [[F2]]
+; NOFUNCSPEC-NEXT:    ret i64 [[ADD]]
+;
+entry:
+  %f1 = call i64 @foo(i64 3, i1 %c1, i1 %c2, i1 %c3, i1 %c4, i1 %c5, i1 %c6, i1 %c7, i1 %c8, i1 %c9, i1 %c10)
+  %f2 = call i64 @foo(i64 4, i1 %c1, i1 %c2, i1 %c3, i1 %c4, i1 %c5, i1 %c6, i1 %c7, i1 %c8, i1 %c9, i1 %c10)
+  %add = add i64 %f1, %f2
+  ret i64 %add
+}
+
+define internal i64 @foo(i64 %n, i1 %c1, i1 %c2, i1 %c3, i1 %c4, i1 %c5, i1 %c6, i1 %c7, i1 %c8, i1 %c9, i1 %c10) {
+entry:
+  br i1 %c1, label %l1, label %l9
+
+l1:
+  %phi1 = phi i64 [ %n, %entry ], [ %phi2, %l2 ]
+  %add = add i64 %phi1, 1
+  %div = sdiv i64 %add, 2
+  br i1 %c2, label %l1_5, label %exit
+
+l1_5:
+  br i1 %c3, label %l1_75, label %l6
+
+l1_75:
+  br i1 %c4, label %l2, label %l3
+
+l2:
+  %phi2 = phi i64 [ %phi1, %l1_75 ], [ %phi3, %l3 ]
+  br label %l1
+
+l3:
+  %phi3 = phi i64 [ %phi1, %l1_75 ], [ %phi4, %l4 ]
+  br label %l2
+
+l4:
+  %phi4 = phi i64 [ %phi5, %l5 ], [ %phi6, %l6 ]
+  br i1 %c5, label %l3, label %l6
+
+l5:
+  %phi5 = phi i64 [ %phi6, %l6_5 ], [ %phi7, %l7 ]
+  br label %l4
+
+l6:
+  %phi6 = phi i64 [ %phi4, %l4 ], [ %phi1, %l1_5 ]
+  br i1 %c6, label %l4, label %l6_5
+
+l6_5:
+  br i1 %c7, label %l5, label %l8
+
+l7:
+  %phi7 = phi i64 [ %phi9, %l9 ], [ %phi8, %l8 ]
+  br i1 %c8, label %l5, label %l8
+
+l8:
+  %phi8 = phi i64 [ %phi6, %l6_5 ], [ %phi7, %l7 ]
+  br i1 %c9, label %l7, label %l9
+
+l9:
+  %phi9 = phi i64 [ %n, %entry ], [ %phi8, %l8 ]
+  %sub = sub i64 %phi9, 1
+  %mul = mul i64 %sub, 2
+  br i1 %c10, label %l7, label %exit
+
+exit:
+  %res = phi i64 [ %div, %l1 ], [ %mul, %l9]
+  ret i64 %res
+}
+