Improve the representation of the atomic builtins in a few ways. First, we make

their call expressions synthetically have the "deduced" types based on their
first argument. We only insert conversions in the AST for arguments whose
values require conversion to match the value type expected. This keeps PR7600
closed by maintaining the return type, but avoids assertions due to unexpected
implicit casts making the type unsigned (test case added from Daniel).

The magic is moved into the codegen for the atomic builtin which inserts the
casts as needed at the IR level to raise the type to an integer suitable for
the LLVM intrinsic. This shouldn't cause any real change in functionality, but
now we can make the builtin be more truly polymorphic.


git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@108638 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 103 insertions, 65 deletions

diff --git a/lib/CodeGen/CGBuiltin.cpp b/lib/CodeGen/CGBuiltin.cpp
index fff4bacab6..0614011932 100644
--- a/lib/CodeGen/CGBuiltin.cpp
+++ b/lib/CodeGen/CGBuiltin.cpp
@@ -41,6 +41,31 @@ static void EmitMemoryBarrier(CodeGenFunction &CGF,
                          C, C + 5);
 }
 
+static Value *EmitCastToInt(CodeGenFunction &CGF,
+                            const llvm::Type *ToType, Value *Val) {
+  if (Val->getType()->isPointerTy()) {
+    return CGF.Builder.CreatePtrToInt(Val, ToType);
+  }
+  assert(Val->getType()->isIntegerTy() &&
+         "Used a non-integer and non-pointer type with atomic builtin");
+  assert(Val->getType()->getScalarSizeInBits() <=
+         ToType->getScalarSizeInBits() && "Integer type too small");
+  return CGF.Builder.CreateSExtOrBitCast(Val, ToType);
+}
+
+static Value *EmitCastFromInt(CodeGenFunction &CGF, QualType ToQualType,
+                              Value *Val) {
+  const llvm::Type *ToType = CGF.ConvertType(ToQualType);
+  if (ToType->isPointerTy()) {
+    return CGF.Builder.CreateIntToPtr(Val, ToType);
+  }
+  assert(Val->getType()->isIntegerTy() &&
+         "Used a non-integer and non-pointer type with atomic builtin");
+  assert(Val->getType()->getScalarSizeInBits() >=
+         ToType->getScalarSizeInBits() && "Integer type too small");
+  return CGF.Builder.CreateTruncOrBitCast(Val, ToType);
+}
+
 // The atomic builtins are also full memory barriers. This is a utility for
 // wrapping a call to the builtins with memory barriers.
 static Value *EmitCallWithBarrier(CodeGenFunction &CGF, Value *Fn,
@@ -60,13 +85,20 @@ static Value *EmitCallWithBarrier(CodeGenFunction &CGF, Value *Fn,
 /// and the expression node.
 static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
                                Intrinsic::ID Id, const CallExpr *E) {
-  Value *Args[2] = { CGF.EmitScalarExpr(E->getArg(0)),
-                     CGF.EmitScalarExpr(E->getArg(1)) };
-  const llvm::Type *ResType[2];
-  ResType[0] = CGF.ConvertType(E->getType());
-  ResType[1] = CGF.ConvertType(E->getArg(0)->getType());
-  Value *AtomF = CGF.CGM.getIntrinsic(Id, ResType, 2);
-  return RValue::get(EmitCallWithBarrier(CGF, AtomF, Args, Args + 2));
+  const llvm::Type *ValueType =
+    llvm::IntegerType::get(CGF.getLLVMContext(),
+                           CGF.getContext().getTypeSize(E->getType()));
+  const llvm::Type *PtrType = ValueType->getPointerTo();
+  const llvm::Type *IntrinsicTypes[2] = { ValueType, PtrType };
+  Value *AtomF = CGF.CGM.getIntrinsic(Id, IntrinsicTypes, 2);
+
+  Value *Args[2] = { CGF.Builder.CreateBitCast(CGF.EmitScalarExpr(E->getArg(0)),
+                                               PtrType),
+                     EmitCastToInt(CGF, ValueType,
+                                   CGF.EmitScalarExpr(E->getArg(1))) };
+  return RValue::get(EmitCastFromInt(CGF, E->getType(),
+                                     EmitCallWithBarrier(CGF, AtomF, Args,
+                                                         Args + 2)));
 }
 
 /// Utility to insert an atomic instruction based Instrinsic::ID and
@@ -75,14 +107,21 @@ static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
                                    Intrinsic::ID Id, const CallExpr *E,
                                    Instruction::BinaryOps Op) {
-  const llvm::Type *ResType[2];
-  ResType[0] = CGF.ConvertType(E->getType());
-  ResType[1] = CGF.ConvertType(E->getArg(0)->getType());
-  Value *AtomF = CGF.CGM.getIntrinsic(Id, ResType, 2);
-  Value *Args[2] = { CGF.EmitScalarExpr(E->getArg(0)),
-                     CGF.EmitScalarExpr(E->getArg(1)) };
+  const llvm::Type *ValueType =
+    llvm::IntegerType::get(CGF.getLLVMContext(),
+                           CGF.getContext().getTypeSize(E->getType()));
+  const llvm::Type *PtrType = ValueType->getPointerTo();
+  const llvm::Type *IntrinsicTypes[2] = { ValueType, PtrType };
+  Value *AtomF = CGF.CGM.getIntrinsic(Id, IntrinsicTypes, 2);
+
+  Value *Args[2] = { CGF.Builder.CreateBitCast(CGF.EmitScalarExpr(E->getArg(0)),
+                                               PtrType),
+                     EmitCastToInt(CGF, ValueType,
+                                   CGF.EmitScalarExpr(E->getArg(1))) };
   Value *Result = EmitCallWithBarrier(CGF, AtomF, Args, Args + 2);
-  return RValue::get(CGF.Builder.CreateBinOp(Op, Result, Args[1]));
+  return RValue::get(EmitCastFromInt(CGF, E->getType(),
+                                     CGF.Builder.CreateBinOp(Op, Result,
+                                                             Args[1])));
 }
 
 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy,
@@ -746,14 +785,23 @@ RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
   case Builtin::BI__sync_val_compare_and_swap_4:
   case Builtin::BI__sync_val_compare_and_swap_8:
   case Builtin::BI__sync_val_compare_and_swap_16: {
-    const llvm::Type *ResType[2];
-    ResType[0]= ConvertType(E->getType());
-    ResType[1] = ConvertType(E->getArg(0)->getType());
-    Value *AtomF = CGM.getIntrinsic(Intrinsic::atomic_cmp_swap, ResType, 2);
-    Value *Args[3] = { EmitScalarExpr(E->getArg(0)),
-                       EmitScalarExpr(E->getArg(1)),
-                       EmitScalarExpr(E->getArg(2)) };
-    return RValue::get(EmitCallWithBarrier(*this, AtomF, Args, Args + 3));
+    const llvm::Type *ValueType =
+      llvm::IntegerType::get(CGF.getLLVMContext(),
+                             CGF.getContext().getTypeSize(E->getType()));
+    const llvm::Type *PtrType = ValueType->getPointerTo();
+    const llvm::Type *IntrinsicTypes[2] = { ValueType, PtrType };
+    Value *AtomF = CGM.getIntrinsic(Intrinsic::atomic_cmp_swap,
+                                    IntrinsicTypes, 2);
+
+    Value *Args[3] = { Builder.CreateBitCast(CGF.EmitScalarExpr(E->getArg(0)),
+                                             PtrType),
+                       EmitCastToInt(CGF, ValueType,
+                                     CGF.EmitScalarExpr(E->getArg(1))),
+                       EmitCastToInt(CGF, ValueType,
+                                     CGF.EmitScalarExpr(E->getArg(2))) };
+    return RValue::get(EmitCastFromInt(CGF, E->getType(),
+                                       EmitCallWithBarrier(CGF, AtomF, Args,
+                                                           Args + 3)));
   }
 
   case Builtin::BI__sync_bool_compare_and_swap_1:
@@ -761,14 +809,22 @@ RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD,
   case Builtin::BI__sync_bool_compare_and_swap_4:
   case Builtin::BI__sync_bool_compare_and_swap_8:
   case Builtin::BI__sync_bool_compare_and_swap_16: {
-    const llvm::Type *ResType[2];
-    ResType[0]= ConvertType(E->getArg(1)->getType());
-    ResType[1] = llvm::PointerType::getUnqual(ResType[0]);
-    Value *AtomF = CGM.getIntrinsic(Intrinsic::atomic_cmp_swap, ResType, 2);
-    Value *OldVal = EmitScalarExpr(E->getArg(1));
-    Value *Args[3] = { EmitScalarExpr(E->getArg(0)),
-                       OldVal,
-                       EmitScalarExpr(E->getArg(2)) };
+    const llvm::Type *ValueType =
+      llvm::IntegerType::get(
+        CGF.getLLVMContext(),
+        CGF.getContext().getTypeSize(E->getArg(1)->getType()));
+    const llvm::Type *PtrType = ValueType->getPointerTo();
+    const llvm::Type *IntrinsicTypes[2] = { ValueType, PtrType };
+    Value *AtomF = CGM.getIntrinsic(Intrinsic::atomic_cmp_swap,
+                                    IntrinsicTypes, 2);
+
+    Value *Args[3] = { Builder.CreateBitCast(CGF.EmitScalarExpr(E->getArg(0)),
+                                             PtrType),
+                       EmitCastToInt(CGF, ValueType,
+                                     CGF.EmitScalarExpr(E->getArg(1))),
+                       EmitCastToInt(CGF, ValueType,
+                                     CGF.EmitScalarExpr(E->getArg(2))) };
+    Value *OldVal = Args[1];
     Value *PrevVal = EmitCallWithBarrier(*this, AtomF, Args, Args + 3);
     Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal);
     // zext bool to int.
diff --git a/lib/Sema/SemaChecking.cpp b/lib/Sema/SemaChecking.cpp
index 370d350329..bedea18203 100644
--- a/lib/Sema/SemaChecking.cpp
+++ b/lib/Sema/SemaChecking.cpp
@@ -512,19 +512,10 @@ Sema::SemaBuiltinAtomicOverloaded(OwningExprResult TheCallResult) {
   FunctionDecl *NewBuiltinDecl =
     cast<FunctionDecl>(LazilyCreateBuiltin(NewBuiltinII, NewBuiltinID,
                                            TUScope, false, DRE->getLocStart()));
-  const FunctionProtoType *BuiltinFT =
-    NewBuiltinDecl->getType()->getAs<FunctionProtoType>();
 
-  QualType OrigValType = ValType;
-  ValType = BuiltinFT->getArgType(0)->getAs<PointerType>()->getPointeeType();
-
-  // If the first type needs to be converted (e.g. void** -> int*), do it now.
-  if (BuiltinFT->getArgType(0) != FirstArg->getType()) {
-    ImpCastExprToType(FirstArg, BuiltinFT->getArgType(0), CastExpr::CK_BitCast);
-    TheCall->setArg(0, FirstArg);
-  }
-
-  // Next, walk the valid ones promoting to the right type.
+  // The first argument is by definition correct, we use it's type as the type
+  // of the entire operation. Walk the remaining arguments promoting them to
+  // the deduced value type.
   for (unsigned i = 0; i != NumFixed; ++i) {
     Expr *Arg = TheCall->getArg(i+1);
 
@@ -564,28 +555,10 @@ Sema::SemaBuiltinAtomicOverloaded(OwningExprResult TheCallResult) {
   UsualUnaryConversions(PromotedCall);
   TheCall->setCallee(PromotedCall);
 
-  // Change the result type of the call to match the result type of the decl.
-  TheCall->setType(NewBuiltinDecl->getCallResultType());
-
-  // If the value type was converted to an integer when processing the
-  // arguments (e.g. void* -> int), we need to convert the result back.
-  if (!Context.hasSameUnqualifiedType(ValType, OrigValType)) {
-    Expr *E = TheCallResult.takeAs<Expr>();
-
-    assert(ValType->isIntegerType() &&
-           "We always convert atomic operation values to integers.");
-    // FIXME: Handle floating point value type here too.
-    CastExpr::CastKind Kind;
-    if (OrigValType->isIntegerType())
-      Kind = CastExpr::CK_IntegralCast;
-    else if (OrigValType->hasPointerRepresentation())
-      Kind = CastExpr::CK_IntegralToPointer;
-    else
-      llvm_unreachable("Unhandled original value type!");
-
-    ImpCastExprToType(E, OrigValType, Kind);
-    return Owned(E);
-  }
+  // Change the result type of the call to match the original value type. This
+  // is arbitrary, but the codegen for these builtins ins design to handle it
+  // gracefully.
+  TheCall->setType(ValType);
 
   return move(TheCallResult);
 }
diff --git a/test/CodeGen/atomic.c b/test/CodeGen/atomic.c
index aa5aa1507b..8b66bfd660 100644
--- a/test/CodeGen/atomic.c
+++ b/test/CodeGen/atomic.c
@@ -1,5 +1,5 @@
 // RUN: %clang_cc1 %s -emit-llvm -o - -triple=i686-apple-darwin9 > %t1
-// RUN: grep @llvm.memory.barrier %t1 | count 38
+// RUN: grep @llvm.memory.barrier %t1 | count 40
 // RUN: grep @llvm.atomic.load.add.i32 %t1 | count 3
 // RUN: grep @llvm.atomic.load.sub.i8 %t1 | count 2
 // RUN: grep @llvm.atomic.load.min.i32 %t1
@@ -19,6 +19,7 @@ int atomic(void)
   int old;
   int val = 1;
   char valc = 1;
+  _Bool valb = 0;
   unsigned int uval = 1;
   int cmp = 0;
 
@@ -43,6 +44,9 @@ int atomic(void)
 
   
   __sync_val_compare_and_swap((void **)0, (void *)0, (void *)0);
+  if ( __sync_val_compare_and_swap(&valb, 0, 1)) {
+    old = 42;
+  }
 
   
   __sync_lock_release(&val);
diff --git a/test/Sema/builtins.c b/test/Sema/builtins.c
index c0a2131868..64a4b30492 100644
--- a/test/Sema/builtins.c
+++ b/test/Sema/builtins.c
@@ -44,6 +44,11 @@ void test9(short v) {
 
   // PR7600: Pointers are implicitly casted to integers and back.
   void *old_ptr = __sync_val_compare_and_swap((void**)0, 0, 0);
+
+  // Ensure the return type is correct even when implicit casts are stripped
+  // away. This triggers an assertion while checking the comparison otherwise.
+  if (__sync_fetch_and_add(&old, 1) == 1) {
+  }
 }