Merge pull request #23950 from eeckstein/fix-compile-time-5.1

[5.1] SILOptimizer: some compile time fixes

Author: eeckstein

include/swift/SILOptimizer/Utils/PerformanceInlinerUtils.h

@@ -312,6 +312,7 @@ class ShortestPathAnalysis {
   SILLoopInfo *LI;
   llvm::DenseMap<const SILBasicBlock *, BlockInfo *> BlockInfos;
   std::vector<BlockInfo> BlockInfoStorage;
+  bool valid = false;
 
   BlockInfo *getBlockInfo(const SILBasicBlock *BB) {
     BlockInfo *BI = BlockInfos[BB];
@@ -381,15 +382,22 @@ class ShortestPathAnalysis {
 public:
   ShortestPathAnalysis(SILFunction *F, SILLoopInfo *LI) : F(F), LI(LI) { }
 
-  bool isValid() const { return !BlockInfos.empty(); }
+  bool isValid() const { return valid; }
 
   /// Compute the distances. The function \p getApplyLength returns the length
   /// of a function call.
   template <typename Func>
   void analyze(ColdBlockInfo &CBI, Func getApplyLength) {
     assert(!isValid());
+    valid = true;
+    unsigned numBlocks = F->size();
 
-    BlockInfoStorage.resize(F->size());
+    // As the complexity of the analysis is more than linear with the number of blocks,
+    // disable it for huge functions. In this case inlining will be less aggressive.
+    if (numBlocks > 2000)
+      return;
+
+    BlockInfoStorage.resize(numBlocks);
 
     // First step: compute the length of the blocks.
     unsigned BlockIdx = 0;
@@ -433,6 +441,11 @@ class ShortestPathAnalysis {
   /// shortest path in the function.
   int getScopeLength(SILBasicBlock *BB, int LoopDepth) {
     assert(BB->getParent() == F);
+
+    // Return a conservative default if the analysis was not done due to a high number of blocks.
+    if (BlockInfos.empty())
+      return ColdBlockLength;
+
     if (LoopDepth >= MaxNumLoopLevels)
       LoopDepth = MaxNumLoopLevels - 1;
     return getBlockInfo(BB)->getScopeLength(LoopDepth);

lib/SILOptimizer/Transforms/SimplifyCFG.cpp

@@ -81,6 +81,10 @@ namespace {
 
     ConstantFolder ConstFolder;
 
+    // True if the function has a large amount of blocks. In this case we turn off some expensive
+    // optimizations.
+    bool isVeryLargeFunction = false;
+
     void constFoldingCallback(SILInstruction *I) {
       // If a terminal instruction gets constant folded (like cond_br), it
       // enables further simplify-CFG optimizations.
@@ -1226,11 +1230,13 @@ bool SimplifyCFG::simplifyBranchBlock(BranchInst *BI) {
       if (DestBB->getArgument(i) != BI->getArg(i)) {
         SILValue Val = BI->getArg(i);
         DestBB->getArgument(i)->replaceAllUsesWith(Val);
-        if (auto *I = dyn_cast<SingleValueInstruction>(Val)) {
-          // Replacing operands may trigger constant folding which then could
-          // trigger other simplify-CFG optimizations.
-          ConstFolder.addToWorklist(I);
-          ConstFolder.processWorkList();
+        if (!isVeryLargeFunction) {
+          if (auto *I = dyn_cast<SingleValueInstruction>(Val)) {
+            // Replacing operands may trigger constant folding which then could
+            // trigger other simplify-CFG optimizations.
+            ConstFolder.addToWorklist(I);
+            ConstFolder.processWorkList();
+          }
         }
       } else {
         // We must be processing an unreachable part of the cfg with a cycle.
@@ -1290,7 +1296,7 @@ bool SimplifyCFG::simplifyBranchBlock(BranchInst *BI) {
   // If this unconditional branch has BBArgs, check to see if duplicating the
   // destination would allow it to be simplified.  This is a simple form of jump
   // threading.
-  if (!BI->getArgs().empty() &&
+  if (!isVeryLargeFunction && !BI->getArgs().empty() &&
       tryJumpThreading(BI))
     return true;
 
@@ -3067,6 +3073,9 @@ bool SimplifyCFG::run() {
 
   LLVM_DEBUG(llvm::dbgs() << "### Run SimplifyCFG on " << Fn.getName() << '\n');
 
+  // Disable some expensive optimizations if the function is huge.
+  isVeryLargeFunction = (Fn.size() > 10000);
+
   // First remove any block not reachable from the entry.
   bool Changed = removeUnreachableBlocks(Fn);
 

lib/SILOptimizer/Utils/PerformanceInlinerUtils.cpp

@@ -418,6 +418,10 @@ ShortestPathAnalysis::Weight ShortestPathAnalysis::
 getWeight(SILBasicBlock *BB, Weight CallerWeight) {
   assert(BB->getParent() == F);
 
+  // Return a conservative default if the analysis was not done due to a high number of blocks.
+  if (BlockInfos.empty())
+    return Weight(CallerWeight.ScopeLength + ColdBlockLength, CallerWeight.LoopWeight);
+
   SILLoop *Loop = LI->getLoopFor(BB);
   if (!Loop) {
     // We are not in a loop. So just account the length of our function scope