[Evaluator] Improve evaluation of call instruction

[evgeny777]

Summary

This allows evaluating calls containing bitcast instruction, which often happens when using ThinLTO.
It's actually a replacement to D43199 and together with D43457 does the same job of optimizing out static ctors, without bothering structure type names.

[grimar]

lib/Transforms/Utils/Evaluator.cpp

243

Seems it better looks like:

Constant *Evaluator::substCallExprArgument(Value *CallExpr, Constant *RV) {
  ConstantExpr *CE = dyn_cast<ConstantExpr>(CallExpr);
  if (!CE || CE->getOpcode() != Instruction::BitCast)
    return RV;

  if (auto *FT = dyn_cast<FunctionType>(CE->getType()->getPointerElementType()))
    return ConstantFoldLoadThroughBitcast(RV, FT->getReturnType(), DL);
  return RV;
}

[evgeny777]

Addressed review comments

[evgeny777]

Any comments on this?

[tejohnson]

Sorry I missed this one.

lib/Transforms/Utils/Evaluator.cpp

232

The name/description seem a little confusing to me. IIUC there isn't any substitution into the function call expression being done here, it is essentially casting the evaluated return value to the call's bitcast type, right?

506

Is this case being tested by the new test case?

test/Transforms/GlobalOpt/evaluate-call.ll

1

Needs comment about what is being tested

[evgeny777]

Thanks for looking at it @tejohnson. Sorry I was too busy to respond earlier.
I've added extra test case to check function calls which can be constant folded.
The substCallExprArgument was changed to castCallResultIfNeeded which might sound more reasonable.

[tejohnson]

LGTM with a couple of nits.

include/llvm/Transforms/Utils/Evaluator.h

78

Document your new functions.

test/Transforms/GlobalOpt/evaluate-constfold-call.ll

2 ↗

(On Diff #152078)

s/ftof/fmodf/?

6 ↗

(On Diff #152078)

Should the resulting code in main() also be checked?

include/llvm/Transforms/Utils/Evaluator.h

Constant *getVal(Value *V) {
     assert(R && "Reference to an uncomputed value!");
     return R;
   }
 
   void setVal(Value *V, Constant *C) {
     ValueStack.back()[V] = C;
   }
 
+  Function *getCallee(Value *V);
+
+  Constant *substCallExprArgument(Value *CallExpr, Constant *RV);
+
   const DenseMap<Constant*, Constant*> &getMutatedMemory() const {
     return MutatedMemory;
   }
 
   const SmallPtrSetImpl<GlobalVariable*> &getInvariants() const {
     return Invariants;
   }
 

lib/Transforms/Utils/Evaluator.cpp

case Instruction::BitCast:
             I, P->getType()->getPointerElementType(), DL);
       break;
     }
   }
 
   return nullptr;  // don't know how to evaluate.
 }
 
+Function *Evaluator::getCallee(Value *V) {
+  auto *CE = dyn_cast<ConstantExpr>(V);
+  if (!CE)
+    return dyn_cast<Function>(getVal(V));
+
+  Constant *C =
+      CE->getOpcode() == Instruction::BitCast
+          ? ConstantFoldLoadThroughBitcast(CE, CE->getOperand(0)->getType(), DL)
+          : CE;
+  return dyn_cast<Function>(C);
+}
+
+/// Subsitute function call in bitcast expression to evaluated return value
+Constant *Evaluator::substCallExprArgument(Value *CallExpr, Constant *RV) {
+  ConstantExpr *CE = dyn_cast<ConstantExpr>(CallExpr);
+  if (!CE || CE->getOpcode() != Instruction::BitCast)
+    return RV;
+
+  if (auto *FT = dyn_cast<FunctionType>(CE->getType()->getPointerElementType())) {
+    RV =  ConstantFoldLoadThroughBitcast(RV, FT->getReturnType(), DL);
+    assert(RV && "Failed to fold bitcast call expr");
+  }
+  return RV;
+}
+
 /// Evaluate all instructions in block BB, returning true if successful, false
 /// if we can't evaluate it.  NewBB returns the next BB that control flows into,
 /// or null upon return.
 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst,
                               BasicBlock *&NextBB) {
   // This is the main evaluation loop.
   while (true) {
     Constant *InstResult = nullptr;
if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) {
           continue;
         }
 
         LLVM_DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n");
         return false;
       }
 
       // Resolve function pointers.
-      Function *Callee = dyn_cast<Function>(getVal(CS.getCalledValue()));
+      Function *Callee = getCallee(CS.getCalledValue());
       if (!Callee || Callee->isInterposable()) {
         LLVM_DEBUG(dbgs() << "Can not resolve function pointer.\n");
         return false;  // Cannot resolve.
       }
 
       SmallVector<Constant*, 8> Formals;
       for (User::op_iterator i = CS.arg_begin(), e = CS.arg_end(); i != e; ++i)
         Formals.push_back(getVal(*i));
 
       if (Callee->isDeclaration()) {
         // If this is a function we can constant fold, do it.
         if (Constant *C = ConstantFoldCall(CS, Callee, Formals, TLI)) {
-          InstResult = C;
+          InstResult = substCallExprArgument(CS.getCalledValue(), C);
           LLVM_DEBUG(dbgs() << "Constant folded function call. Result: "
                             << *InstResult << "\n");
         } else {
           LLVM_DEBUG(dbgs() << "Can not constant fold function call.\n");
           return false;
         }
       } else {
         if (Callee->getFunctionType()->isVarArg()) {
           LLVM_DEBUG(dbgs() << "Can not constant fold vararg function call.\n");
           return false;
         }
 
         Constant *RetVal = nullptr;
         // Execute the call, if successful, use the return value.
         ValueStack.emplace_back();
         if (!EvaluateFunction(Callee, RetVal, Formals)) {
           LLVM_DEBUG(dbgs() << "Failed to evaluate function.\n");
           return false;
         }
         ValueStack.pop_back();
-        InstResult = RetVal;
+        InstResult = substCallExprArgument(CS.getCalledValue(), RetVal);
 
         if (InstResult) {
           LLVM_DEBUG(dbgs() << "Successfully evaluated function. Result: "
                             << *InstResult << "\n\n");
         } else {
           LLVM_DEBUG(dbgs()
                      << "Successfully evaluated function. Result: 0\n\n");
         }

test/Transforms/GlobalOpt/evaluate-call.ll

+; RUN: opt -globalopt -instcombine %s -S -o - | FileCheck %s
+
+; CHECK:      i32 @main()
+; CHECK-NEXT:   %1 = load i64, i64* inttoptr (i64 32 to i64*), align 32
+; CHECK-NEXT:   %2 = trunc i64 %1 to i32
+; CHECK-NEXT:   ret i32 %2
+; CHECK-NOT: _GLOBAL__sub_I_main
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc-linux-gnu"
+
+%struct.Bar = type { i64, i64 }
+%struct.S = type { %struct.Foo* }
+%struct.Foo = type { i64, i64 }
+%struct.Baz = type { i64, i64, %struct.Bar }
+
+@instance = internal local_unnamed_addr global %struct.S zeroinitializer, align 8
[e-mail]_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @_GLOBAL__sub_I_main.cpp, i8* null }]
+@gBaz = available_externally dso_local local_unnamed_addr global %struct.Baz* null, align 8
+@gFoo = available_externally dso_local local_unnamed_addr global %struct.Bar* null, align 8
+
+; Function Attrs: norecurse nounwind readonly uwtable
+define dso_local i32 @main() local_unnamed_addr {
+  %1 = load %struct.Foo*, %struct.Foo** getelementptr inbounds (%struct.S, %struct.S* @instance, i64 0, i32 0), align 8
+  %2 = getelementptr inbounds %struct.Foo, %struct.Foo* %1, i64 0, i32 0
+  %3 = load i64, i64* %2, align 8
+  %4 = trunc i64 %3 to i32
+  ret i32 %4
+}
+
+; Function Attrs: uwtable
+define internal void @_GLOBAL__sub_I_main.cpp() section ".text.startup" {
+  %1 = tail call %struct.Foo* bitcast (%struct.Bar* ()* @_Z6getFoov to %struct.Foo* ()*)()
+  %2 = getelementptr inbounds %struct.Foo, %struct.Foo* %1, i64 1
+  store %struct.Foo* %2, %struct.Foo** getelementptr inbounds (%struct.S, %struct.S* @instance, i64 0, i32 0), align 8
+  ret void
+}
+
+; Function Attrs: norecurse nounwind readonly uwtable
+define available_externally dso_local %struct.Bar* @_Z6getFoov() local_unnamed_addr {
+  %1 = load %struct.Bar*, %struct.Bar** @gFoo, align 8
+  %2 = icmp eq %struct.Bar* %1, null
+  %3 = load %struct.Baz*, %struct.Baz** @gBaz, align 8
+  %4 = getelementptr inbounds %struct.Baz, %struct.Baz* %3, i64 0, i32 2
+  %5 = select i1 %2, %struct.Bar* %4, %struct.Bar* %1
+  ret %struct.Bar* %5
+}