Add support for simplifying a load from a computed value to a load from a global...

[oota-llvm.git] / lib / Analysis / ConstantFolding.cpp

diff --git a/lib/Analysis/ConstantFolding.cpp b/lib/Analysis/ConstantFolding.cpp
index 0ac512726ac18fec2afbbaa952bd6da5d0d7355a..0bf7967e83b132679b81105e21951703606a571c 100644 (file)
--- a/lib/Analysis/ConstantFolding.cpp
+++ b/lib/Analysis/ConstantFolding.cpp
@@ -208,7 +208,7 @@ static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
          i != e; ++i, ++GTI) {
       ConstantInt *CI = dyn_cast<ConstantInt>(*i);
       if (!CI) return false;  // Index isn't a simple constant?
-      if (CI->getZExtValue() == 0) continue;  // Not adding anything.
+      if (CI->isZero()) continue;  // Not adding anything.
       
       if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
         // N = N + Offset
@@ -401,7 +401,7 @@ static Constant *FoldReinterpretLoadFromConstPtr(Constant *C,
   APInt ResultVal = APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1]);
   for (unsigned i = 1; i != BytesLoaded; ++i) {
     ResultVal <<= 8;
-    ResultVal |= APInt(IntType->getBitWidth(), RawBytes[BytesLoaded-1-i]);
+    ResultVal |= RawBytes[BytesLoaded-1-i];
   }
 
   return ConstantInt::get(IntType->getContext(), ResultVal);
@@ -436,8 +436,10 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
     unsigned StrLen = Str.length();
     const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
     unsigned NumBits = Ty->getPrimitiveSizeInBits();
-    // Replace LI with immediate integer store.
-    if ((NumBits >> 3) == StrLen + 1) {
+    // Replace load with immediate integer if the result is an integer or fp
+    // value.
+    if ((NumBits >> 3) == StrLen + 1 && (NumBits & 7) == 0 &&
+        (isa<IntegerType>(Ty) || Ty->isFloatingPointTy())) {
       APInt StrVal(NumBits, 0);
       APInt SingleChar(NumBits, 0);
       if (TD->isLittleEndian()) {
@@ -454,7 +456,11 @@ Constant *llvm::ConstantFoldLoadFromConstPtr(Constant *C,
         SingleChar = 0;
         StrVal = (StrVal << 8) | SingleChar;
       }
-      return ConstantInt::get(CE->getContext(), StrVal);
+      
+      Constant *Res = ConstantInt::get(CE->getContext(), StrVal);
+      if (Ty->isFloatingPointTy())
+        Res = ConstantExpr::getBitCast(Res, Ty);
+      return Res;
     }
   }
   
@@ -772,9 +778,9 @@ Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
   case Instruction::ICmp:
   case Instruction::FCmp: assert(0 && "Invalid for compares");
   case Instruction::Call:
-    if (Function *F = dyn_cast<Function>(Ops[0]))
+    if (Function *F = dyn_cast<Function>(Ops[NumOps - 1]))
       if (canConstantFoldCallTo(F))
-        return ConstantFoldCall(F, Ops+1, NumOps-1);
+        return ConstantFoldCall(F, Ops, NumOps - 1);
     return 0;
   case Instruction::PtrToInt:
     // If the input is a inttoptr, eliminate the pair.  This requires knowing
@@ -994,6 +1000,8 @@ llvm::canConstantFoldCallTo(const Function *F) {
   case Intrinsic::usub_with_overflow:
   case Intrinsic::sadd_with_overflow:
   case Intrinsic::ssub_with_overflow:
+  case Intrinsic::convert_from_fp16:
+  case Intrinsic::convert_to_fp16:
     return true;
   default:
     return false;
@@ -1074,6 +1082,15 @@ llvm::ConstantFoldCall(Function *F,
   const Type *Ty = F->getReturnType();
   if (NumOperands == 1) {
     if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
+      if (Name == "llvm.convert.to.fp16") {
+        APFloat Val(Op->getValueAPF());
+
+        bool lost = false;
+        Val.convert(APFloat::IEEEhalf, APFloat::rmNearestTiesToEven, &lost);
+
+        return ConstantInt::get(F->getContext(), Val.bitcastToAPInt());
+      }
+
       if (!Ty->isFloatTy() && !Ty->isDoubleTy())
         return 0;
       /// Currently APFloat versions of these functions do not exist, so we use
@@ -1158,6 +1175,20 @@ llvm::ConstantFoldCall(Function *F,
         return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
       else if (Name.startswith("llvm.ctlz"))
         return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
+      else if (Name == "llvm.convert.from.fp16") {
+        APFloat Val(Op->getValue());
+
+        bool lost = false;
+        APFloat::opStatus status =
+          Val.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &lost);
+
+        // Conversion is always precise.
+        status = status;
+        assert(status == APFloat::opOK && !lost &&
+               "Precision lost during fp16 constfolding");
+
+        return ConstantFP::get(F->getContext(), Val);
+      }
       return 0;
     }