fp_round's produced by getCopyFromParts should always be exact, because

they are produced by calls (which are known exact) and by cross block copies
which are known to be produced by extends.

This improves:

define double @test2() {
	%tmp85 = call double asm sideeffect "fld0", "={st(0)}"()
	ret double %tmp85
}

from:

_test2:
	subl	$20, %esp
	# InlineAsm Start
	fld0
	# InlineAsm End
	fstpl	8(%esp)
	movsd	8(%esp), %xmm0
	movsd	%xmm0, (%esp)
	fldl	(%esp)
	addl	$20, %esp
	#FP_REG_KILL
	ret

to:

_test2:
	# InlineAsm Start
	fld0
	# InlineAsm End
	#FP_REG_KILL
	ret

by avoiding a f64 <-> f80 trip

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@48108 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index 97e12a4a0e9a84bd0705034d43e1c23928353788..94386c62851f485f9a4bafc5853147e1b6b40b9f 100644 (file)
--- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -628,15 +628,13 @@ public:
 /// combined into the value they represent.  If the parts combine to a type
 /// larger then ValueVT then AssertOp can be used to specify whether the extra
 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
-/// (ISD::AssertSext).  Likewise TruncExact is used for floating point types to
-/// indicate that the extra bits can be discarded without losing precision.
+/// (ISD::AssertSext).
 static SDOperand getCopyFromParts(SelectionDAG &DAG,
                                   const SDOperand *Parts,
                                   unsigned NumParts,
                                   MVT::ValueType PartVT,
                                   MVT::ValueType ValueVT,
-                                  ISD::NodeType AssertOp = ISD::DELETED_NODE,
-                                  bool TruncExact = false) {
+                                  ISD::NodeType AssertOp = ISD::DELETED_NODE) {
   assert(NumParts > 0 && "No parts to assemble!");
   TargetLowering &TLI = DAG.getTargetLoweringInfo();
   SDOperand Val = Parts[0];
@@ -761,8 +759,9 @@ static SDOperand getCopyFromParts(SelectionDAG &DAG,
 
   if (MVT::isFloatingPoint(PartVT) && MVT::isFloatingPoint(ValueVT)) {
     if (ValueVT < Val.getValueType())
+      // FP_ROUND's are always exact here.
       return DAG.getNode(ISD::FP_ROUND, ValueVT, Val,
-                         DAG.getIntPtrConstant(TruncExact));
+                         DAG.getIntPtrConstant(1));
     return DAG.getNode(ISD::FP_EXTEND, ValueVT, Val);
   }
 
@@ -3267,7 +3266,8 @@ SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
   }
   
   // Assemble the legal parts into the final value.
-  return getCopyFromParts(DAG, &Parts[0], NumParts, RegVT, ValueVT);
+  return getCopyFromParts(DAG, &Parts[0], NumParts, RegVT, ValueVT,
+                          ISD::DELETED_NODE);
 }
 
 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
@@ -4147,7 +4147,7 @@ TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
       AssertOp = ISD::AssertZext;
 
     Ops.push_back(getCopyFromParts(DAG, &Parts[0], NumParts, PartVT, VT,
-                                   AssertOp, true));
+                                   AssertOp));
   }
   assert(i == NumArgRegs && "Argument register count mismatch!");
   return Ops;
@@ -4257,7 +4257,7 @@ TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
     for (unsigned i = 0; i != NumRegs; ++i)
       Results[i] = Res.getValue(i);
     Res = getCopyFromParts(DAG, &Results[0], NumRegs, RegisterVT, VT,
-                           AssertOp, true);
+                           AssertOp);
   }
 
   return std::make_pair(Res, Chain);