Preserve the assumption cache more often

We were clearing it out in LoopUnswitch and InlineFunction instead of
attempting to preserve it.

llvm-svn: 278860

lib/Transforms/Scalar/LoopUnswitch.cpp

@@ -1078,10 +1078,6 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
                                 F->getBasicBlockList(),
                                 NewBlocks[0]->getIterator(), F->end());
 
-  // FIXME: We could register any cloned assumptions instead of clearing the
-  // whole function's cache.
-  AC->clear();
-
   // Now we create the new Loop object for the versioned loop.
   Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
 
@@ -1131,10 +1127,15 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
   }
 
   // Rewrite the code to refer to itself.
-  for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
+  for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) {
-    for (Instruction &I : *NewBlocks[i])
+    for (Instruction &I : *NewBlocks[i]) {
       RemapInstruction(&I, VMap,
                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
+      if (auto *II = dyn_cast<IntrinsicInst>(&I))
+        if (II->getIntrinsicID() == Intrinsic::assume)
+          AC->registerAssumption(II);
+    }
+  }
 
   // Rewrite the original preheader to select between versions of the loop.
   BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());

lib/Transforms/Utils/InlineFunction.cpp

@@ -1055,6 +1055,9 @@ static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
   if (!PreserveAlignmentAssumptions)
     return;
+  AssumptionCache *AC = IFI.GetAssumptionCache
+                            ? &(*IFI.GetAssumptionCache)(*CS.getCaller())
+                            : nullptr;
   auto &DL = CS.getCaller()->getParent()->getDataLayout();
 
   // To avoid inserting redundant assumptions, we should check for assumptions
@@ -1077,13 +1080,13 @@ static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
       // If we can already prove the asserted alignment in the context of the
       // caller, then don't bother inserting the assumption.
       Value *Arg = CS.getArgument(I->getArgNo());
-      if (getKnownAlignment(Arg, DL, CS.getInstruction(),
+      if (getKnownAlignment(Arg, DL, CS.getInstruction(), AC, &DT) >= Align)
-                            &(*IFI.GetAssumptionCache)(*CS.getCaller()),
-                            &DT) >= Align)
         continue;
 
-      IRBuilder<>(CS.getInstruction())
+      CallInst *NewAssumption = IRBuilder<>(CS.getInstruction())
-          .CreateAlignmentAssumption(DL, Arg, Align);
+                                    .CreateAlignmentAssumption(DL, Arg, Align);
+      if (AC)
+        AC->registerAssumption(NewAssumption);
     }
   }
 }
@@ -1194,12 +1197,13 @@ static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
     if (ByValAlignment <= 1)  // 0 = unspecified, 1 = no particular alignment.
       return Arg;
 
+    AssumptionCache *AC =
+        IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr;
     const DataLayout &DL = Caller->getParent()->getDataLayout();
 
     // If the pointer is already known to be sufficiently aligned, or if we can
     // round it up to a larger alignment, then we don't need a temporary.
-    if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall,
+    if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall, AC) >=
-                                   &(*IFI.GetAssumptionCache)(*Caller)) >=
         ByValAlignment)
       return Arg;
     
@@ -1609,10 +1613,15 @@ bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
     // Propagate llvm.mem.parallel_loop_access if necessary.
     PropagateParallelLoopAccessMetadata(CS, VMap);
 
-    // FIXME: We could register any cloned assumptions instead of clearing the
+    // Register any cloned assumptions.
-    // whole function's cache.
     if (IFI.GetAssumptionCache)
-      (*IFI.GetAssumptionCache)(*Caller).clear();
+      for (BasicBlock &NewBlock :
+           make_range(FirstNewBlock->getIterator(), Caller->end()))
+        for (Instruction &I : NewBlock) {
+          if (auto *II = dyn_cast<IntrinsicInst>(&I))
+            if (II->getIntrinsicID() == Intrinsic::assume)
+              (*IFI.GetAssumptionCache)(*Caller).registerAssumption(II);
+        }
   }
 
   // If there are any alloca instructions in the block that used to be the entry
@@ -2130,9 +2139,10 @@ bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
   // the entries are the same or undef).  If so, remove the PHI so it doesn't
   // block other optimizations.
   if (PHI) {
+    AssumptionCache *AC =
+        IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr;
     auto &DL = Caller->getParent()->getDataLayout();
-    if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr,
+    if (Value *V = SimplifyInstruction(PHI, DL, nullptr, nullptr, AC)) {
-                                       &(*IFI.GetAssumptionCache)(*Caller))) {
       PHI->replaceAllUsesWith(V);
       PHI->eraseFromParent();
     }