Add Neon VCVT instructions for f32 <-> f16 conversions.

Clang is now providing intrinsics for these and so we need to support them
in the backend.  Radar 8068427.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@121902 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 22 insertions, 1 deletions

diff --git a/lib/Target/ARM/ARM.td b/lib/Target/ARM/ARM.td
index 5b43716..bf4315f 100644
--- a/lib/Target/ARM/ARM.td
+++ b/lib/Target/ARM/ARM.td
@@ -103,7 +103,8 @@ def ProcA8      : SubtargetFeature<"a8", "ARMProcFamily", "CortexA8",
                                     FeatureHasSlowFPVMLx, FeatureT2XtPk]>;
 def ProcA9      : SubtargetFeature<"a9", "ARMProcFamily", "CortexA9",
                                    "Cortex-A9 ARM processors",
-                                   [FeatureHasSlowFPVMLx, FeatureT2XtPk]>;
+                                   [FeatureHasSlowFPVMLx, FeatureT2XtPk,
+                                    FeatureFP16]>;
 
 class ProcNoItin<string Name, list<SubtargetFeature> Features>
  : Processor<Name, GenericItineraries, Features>;
diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index 71948e1..1686b80 100644
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -156,6 +156,7 @@ def NoVFP            : Predicate<"!Subtarget->hasVFP2()">;
 def HasVFP2          : Predicate<"Subtarget->hasVFP2()">, AssemblerPredicate;
 def HasVFP3          : Predicate<"Subtarget->hasVFP3()">, AssemblerPredicate;
 def HasNEON          : Predicate<"Subtarget->hasNEON()">, AssemblerPredicate;
+def HasFP16          : Predicate<"Subtarget->hasFP16()">, AssemblerPredicate;
 def HasDivide        : Predicate<"Subtarget->hasDivide()">, AssemblerPredicate;
 def HasT2ExtractPack : Predicate<"Subtarget->hasT2ExtractPack()">,
                                  AssemblerPredicate;
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index b95e02f..e27a4a8 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -1724,6 +1724,15 @@ class N2VL<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
         (ins DPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
         [(set QPR:$Vd, (TyQ (OpNode (TyD DPR:$Vm))))]>;
 
+// Long 2-register intrinsics.
+class N2VLInt<bits<2> op24_23, bits<2> op21_20, bits<2> op19_18,
+              bits<2> op17_16, bits<5> op11_7, bit op6, bit op4,
+              InstrItinClass itin, string OpcodeStr, string Dt,
+              ValueType TyQ, ValueType TyD, Intrinsic IntOp>
+  : N2V<op24_23, op21_20, op19_18, op17_16, op11_7, op6, op4, (outs QPR:$Vd),
+        (ins DPR:$Vm), itin, OpcodeStr, Dt, "$Vd, $Vm", "",
+        [(set QPR:$Vd, (TyQ (IntOp (TyD DPR:$Vm))))]>;
+
 // 2-register shuffles (VTRN/VZIP/VUZP), both double- and quad-register.
 class N2VDShuffle<bits<2> op19_18, bits<5> op11_7, string OpcodeStr, string Dt>
   : N2V<0b11, 0b11, op19_18, 0b10, op11_7, 0, 0, (outs DPR:$Vd, DPR:$Vm),
@@ -4447,6 +4456,16 @@ def VCVTxs2fq : N2VCvtQ<0, 1, 0b1110, 0, 1, "vcvt", "f32.s32",
 def VCVTxu2fq : N2VCvtQ<1, 1, 0b1110, 0, 1, "vcvt", "f32.u32",
                         v4f32, v4i32, int_arm_neon_vcvtfxu2fp>;
 
+//   VCVT     : Vector Convert Between Half-Precision and Single-Precision.
+def  VCVTf2h  : N2VNInt<0b11, 0b11, 0b01, 0b10, 0b01100, 0, 0,
+                        IIC_VUNAQ, "vcvt", "f16.f32",
+                        v4i16, v4f32, int_arm_neon_vcvtfp2hf>,
+                Requires<[HasNEON, HasFP16]>;
+def  VCVTh2f  : N2VLInt<0b11, 0b11, 0b01, 0b10, 0b01110, 0, 0,
+                        IIC_VUNAQ, "vcvt", "f32.f16",
+                        v4f32, v4i16, int_arm_neon_vcvthf2fp>,
+                Requires<[HasNEON, HasFP16]>;
+
 // Vector Reverse.
 
 //   VREV64   : Vector Reverse elements within 64-bit doublewords