Mark FP_EXTEND form v2f32 to v2f64 as "expand" for ARM NEON. Patch by Pete Couperus.
authorEli Friedman <eli.friedman@gmail.com>
Sat, 17 Nov 2012 01:52:46 +0000 (01:52 +0000)
committerEli Friedman <eli.friedman@gmail.com>
Sat, 17 Nov 2012 01:52:46 +0000 (01:52 +0000)
llvm-svn: 168240

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
llvm/lib/Target/ARM/ARMISelLowering.cpp
llvm/test/CodeGen/ARM/neon_fpconv.ll

index 3250133..d63862d 100644 (file)
@@ -222,6 +222,7 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
   case ISD::FNEARBYINT:
   case ISD::FFLOOR:
   case ISD::FP_ROUND:
+  case ISD::FP_EXTEND:
   case ISD::FMA:
   case ISD::SIGN_EXTEND_INREG:
     QueryType = Node->getValueType(0);
index e123f15..bd470eb 100644 (file)
@@ -544,6 +544,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
 
     setOperationAction(ISD::FP_ROUND,   MVT::v2f32, Expand);
+    setOperationAction(ISD::FP_EXTEND,  MVT::v2f64, Expand);
 
     setTargetDAGCombine(ISD::INTRINSIC_VOID);
     setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
index f80ea3e..1948ad8 100644 (file)
@@ -7,3 +7,11 @@ define <2 x float> @vtrunc(<2 x double> %a) {
   %vt = fptrunc <2 x double> %a to <2 x float>
   ret <2 x float> %vt
 }
+
+define <2 x double> @vextend(<2 x float> %a) {
+; CHECK: vcvt.f64.f32 [[D0:d[0-9]+]], [[S0:s[0-9]+]]
+; CHECK: vcvt.f64.f32 [[D1:d[0-9]+]], [[S1:s[0-9]+]]
+  %ve = fpext <2 x float> %a to <2 x double>
+  ret <2 x double> %ve
+}
+