Add the Int64X2# primitive type and associated primops.

5 files changed, 149 insertions, 3 deletions

diff --git a/compiler/codeGen/StgCmmPrim.hs b/compiler/codeGen/StgCmmPrim.hs
index f443fdebbe..10a514b6e1 100644
--- a/compiler/codeGen/StgCmmPrim.hs
+++ b/compiler/codeGen/StgCmmPrim.hs
@@ -387,6 +387,8 @@ emitPrimOp _      res IndexOffAddrOp_DoubleX2         args = doIndexOffAddrOp
 emitPrimOp _      res IndexOffAddrOp_DoubleAsDoubleX2 args = doIndexOffAddrOpAs Nothing vec2f64 f64 res args
 emitPrimOp _      res IndexOffAddrOp_Int32X4          args = doIndexOffAddrOp   Nothing vec4b32 res args
 emitPrimOp _      res IndexOffAddrOp_Int32AsInt32X4   args = doIndexOffAddrOpAs Nothing vec4b32 b32 res args
+emitPrimOp _      res IndexOffAddrOp_Int64X2          args = doIndexOffAddrOp   Nothing vec2b64 res args
+emitPrimOp _      res IndexOffAddrOp_Int64AsInt64X2   args = doIndexOffAddrOpAs Nothing vec2b64 b64 res args
 
 -- ReadXXXoffAddr, which are identical, for our purposes, to IndexXXXoffAddr.
 
@@ -412,6 +414,8 @@ emitPrimOp _      res ReadOffAddrOp_DoubleX2         args = doIndexOffAddrOp   N
 emitPrimOp _      res ReadOffAddrOp_DoubleAsDoubleX2 args = doIndexOffAddrOpAs Nothing vec2f64 b64 res args
 emitPrimOp _      res ReadOffAddrOp_Int32X4          args = doIndexOffAddrOp   Nothing vec4b32 res args
 emitPrimOp _      res ReadOffAddrOp_Int32AsInt32X4   args = doIndexOffAddrOpAs Nothing vec4b32 b32 res args
+emitPrimOp _      res ReadOffAddrOp_Int64X2          args = doIndexOffAddrOp   Nothing vec2b64 res args
+emitPrimOp _      res ReadOffAddrOp_Int64AsInt64X2   args = doIndexOffAddrOpAs Nothing vec2b64 b64 res args
 
 -- IndexXXXArray
 
@@ -437,6 +441,8 @@ emitPrimOp _      res IndexByteArrayOp_DoubleX2         args = doIndexByteArrayO
 emitPrimOp _      res IndexByteArrayOp_DoubleAsDoubleX2 args = doIndexByteArrayOpAs Nothing vec2f64 f64 res args
 emitPrimOp _      res IndexByteArrayOp_Int32X4          args = doIndexByteArrayOp   Nothing vec4b32 res args
 emitPrimOp _      res IndexByteArrayOp_Int32AsInt32X4   args = doIndexByteArrayOpAs Nothing vec4b32 b32 res args
+emitPrimOp _      res IndexByteArrayOp_Int64X2          args = doIndexByteArrayOp   Nothing vec2b64 res args
+emitPrimOp _      res IndexByteArrayOp_Int64AsInt64X2   args = doIndexByteArrayOpAs Nothing vec2b64 b64 res args
 
 -- ReadXXXArray, identical to IndexXXXArray.
 
@@ -462,6 +468,8 @@ emitPrimOp _      res ReadByteArrayOp_DoubleX2         args = doIndexByteArrayOp
 emitPrimOp _      res ReadByteArrayOp_DoubleAsDoubleX2 args = doIndexByteArrayOpAs Nothing vec2f64 f64 res args
 emitPrimOp _      res ReadByteArrayOp_Int32X4          args = doIndexByteArrayOp   Nothing vec4b32 res args
 emitPrimOp _      res ReadByteArrayOp_Int32AsInt32X4   args = doIndexByteArrayOpAs Nothing vec4b32 b32 res args
+emitPrimOp _      res ReadByteArrayOp_Int64X2          args = doIndexByteArrayOp   Nothing vec2b64 res args
+emitPrimOp _      res ReadByteArrayOp_Int64AsInt64X2   args = doIndexByteArrayOpAs Nothing vec2b64 b64 res args
 
 -- WriteXXXoffAddr
 
@@ -487,6 +495,8 @@ emitPrimOp _      res WriteOffAddrOp_DoubleX2         args = doWriteOffAddrOp No
 emitPrimOp _      res WriteOffAddrOp_DoubleAsDoubleX2 args = doWriteOffAddrOp Nothing f64 res args
 emitPrimOp _      res WriteOffAddrOp_Int32X4          args = doWriteOffAddrOp Nothing vec4b32 res args
 emitPrimOp _      res WriteOffAddrOp_Int32AsInt32X4   args = doWriteOffAddrOp Nothing b32 res args
+emitPrimOp _      res WriteOffAddrOp_Int64X2          args = doWriteOffAddrOp Nothing vec2b64 res args
+emitPrimOp _      res WriteOffAddrOp_Int64AsInt64X2   args = doWriteOffAddrOp Nothing b64 res args
 
 -- WriteXXXArray
 
@@ -512,6 +522,8 @@ emitPrimOp _      res WriteByteArrayOp_DoubleX2         args = doWriteByteArrayO
 emitPrimOp _      res WriteByteArrayOp_DoubleAsDoubleX2 args = doWriteByteArrayOp Nothing f64 res args
 emitPrimOp _      res WriteByteArrayOp_Int32X4          args = doWriteByteArrayOp Nothing vec4b32 res args
 emitPrimOp _      res WriteByteArrayOp_Int32AsInt32X4   args = doWriteByteArrayOp Nothing b32 res args
+emitPrimOp _      res WriteByteArrayOp_Int64X2          args = doWriteByteArrayOp Nothing vec2b64 res args
+emitPrimOp _      res WriteByteArrayOp_Int64AsInt64X2   args = doWriteByteArrayOp Nothing b64 res args
 
 -- Copying and setting byte arrays
 emitPrimOp _      [] CopyByteArrayOp [src,src_off,dst,dst_off,n] =
@@ -589,6 +601,24 @@ emitPrimOp dflags res@[_,_,_,_] Int32X4UnpackOp [arg] =
 emitPrimOp dflags [res] Int32X4InsertOp [v,e,i] =
     doVecInsertOp (Just (mo_WordTo32 dflags)) vec4b32 v e i res
 
+emitPrimOp _ [res] Int64ToInt64X2Op [e] =
+    doVecPackOp Nothing vec2b64 zero [e,e] res
+  where
+    zero :: CmmExpr
+    zero = CmmLit $ CmmVec (replicate 2 (CmmInt 0 W64))
+
+emitPrimOp _ [res] Int64X2PackOp es@[_,_] =
+    doVecPackOp Nothing vec2b64 zero es res
+  where
+    zero :: CmmExpr
+    zero = CmmLit $ CmmVec (replicate 2 (CmmInt 0 W64))
+
+emitPrimOp _ res@[_,_] Int64X2UnpackOp [arg] =
+    doVecUnpackOp Nothing vec2b64 arg res
+
+emitPrimOp _ [res] Int64X2InsertOp [v,e,i] =
+    doVecInsertOp Nothing vec2b64 v e i res
+
 -- The rest just translate straightforwardly
 emitPrimOp dflags [res] op [arg]
    | nopOp op
@@ -916,6 +946,13 @@ translateOp _ Int32X4QuotOp  = Just (MO_VS_Quot 4 W32)
 translateOp _ Int32X4RemOp   = Just (MO_VS_Rem  4 W32)
 translateOp _ Int32X4NegOp   = Just (MO_VS_Neg  4 W32)
 
+translateOp _ Int64X2AddOp   = Just (MO_V_Add   2 W64)
+translateOp _ Int64X2SubOp   = Just (MO_V_Sub   2 W64)
+translateOp _ Int64X2MulOp   = Just (MO_V_Mul   2 W64)
+translateOp _ Int64X2QuotOp  = Just (MO_VS_Quot 2 W64)
+translateOp _ Int64X2RemOp   = Just (MO_VS_Rem  2 W64)
+translateOp _ Int64X2NegOp   = Just (MO_VS_Neg  2 W64)
+
 -- Conversions
 
 translateOp dflags Int2DoubleOp   = Just (MO_SF_Conv (wordWidth dflags) W64)
diff --git a/compiler/prelude/PrelNames.lhs b/compiler/prelude/PrelNames.lhs
index 3e47d9d79c..03a95dee8f 100644
--- a/compiler/prelude/PrelNames.lhs
+++ b/compiler/prelude/PrelNames.lhs
@@ -1421,11 +1421,13 @@ typeNatMulTyFamNameKey    = mkPreludeTyConUnique 163
 typeNatExpTyFamNameKey    = mkPreludeTyConUnique 164
 
 -- SIMD vector types (Unique keys)
-floatX4PrimTyConKey, doubleX2PrimTyConKey, int32X4PrimTyConKey :: Unique
+floatX4PrimTyConKey, doubleX2PrimTyConKey, int32X4PrimTyConKey,
+  int64X2PrimTyConKey :: Unique
 
 floatX4PrimTyConKey  = mkPreludeTyConUnique 170
 doubleX2PrimTyConKey = mkPreludeTyConUnique 171
 int32X4PrimTyConKey  = mkPreludeTyConUnique 172
+int64X2PrimTyConKey  = mkPreludeTyConUnique 173
 
 ---------------- Template Haskell -------------------
 --      USES TyConUniques 200-299
diff --git a/compiler/prelude/TysPrim.lhs b/compiler/prelude/TysPrim.lhs
index cf7a2666b4..c59884ba33 100644
--- a/compiler/prelude/TysPrim.lhs
+++ b/compiler/prelude/TysPrim.lhs
@@ -78,7 +78,8 @@ module TysPrim(
         -- * SIMD
 	floatX4PrimTyCon,		floatX4PrimTy,
 	doubleX2PrimTyCon,		doubleX2PrimTy,
-	int32X4PrimTyCon,		int32X4PrimTy
+	int32X4PrimTyCon,		int32X4PrimTy,
+	int64X2PrimTyCon,		int64X2PrimTy
   ) where
 
 #include "HsVersions.h"
@@ -144,6 +145,7 @@ primTyCons
     , floatX4PrimTyCon
     , doubleX2PrimTyCon
     , int32X4PrimTyCon
+    , int64X2PrimTyCon
     ]
 
 mkPrimTc :: FastString -> Unique -> TyCon -> Name
@@ -153,7 +155,7 @@ mkPrimTc fs unique tycon
 		  (ATyCon tycon)	-- Relevant TyCon
 		  UserSyntax		-- None are built-in syntax
 
-charPrimTyConName, intPrimTyConName, int32PrimTyConName, int64PrimTyConName, wordPrimTyConName, word32PrimTyConName, word64PrimTyConName, addrPrimTyConName, floatPrimTyConName, doublePrimTyConName, statePrimTyConName, realWorldTyConName, arrayPrimTyConName, arrayArrayPrimTyConName, byteArrayPrimTyConName, mutableArrayPrimTyConName, mutableByteArrayPrimTyConName, mutableArrayArrayPrimTyConName, mutVarPrimTyConName, mVarPrimTyConName, tVarPrimTyConName, stablePtrPrimTyConName, stableNamePrimTyConName, bcoPrimTyConName, weakPrimTyConName, threadIdPrimTyConName, eqPrimTyConName, floatX4PrimTyConName, doubleX2PrimTyConName, int32X4PrimTyConName :: Name
+charPrimTyConName, intPrimTyConName, int32PrimTyConName, int64PrimTyConName, wordPrimTyConName, word32PrimTyConName, word64PrimTyConName, addrPrimTyConName, floatPrimTyConName, doublePrimTyConName, statePrimTyConName, realWorldTyConName, arrayPrimTyConName, arrayArrayPrimTyConName, byteArrayPrimTyConName, mutableArrayPrimTyConName, mutableByteArrayPrimTyConName, mutableArrayArrayPrimTyConName, mutVarPrimTyConName, mVarPrimTyConName, tVarPrimTyConName, stablePtrPrimTyConName, stableNamePrimTyConName, bcoPrimTyConName, weakPrimTyConName, threadIdPrimTyConName, eqPrimTyConName, floatX4PrimTyConName, doubleX2PrimTyConName, int32X4PrimTyConName, int64X2PrimTyConName :: Name
 charPrimTyConName    	      = mkPrimTc (fsLit "Char#") charPrimTyConKey charPrimTyCon
 intPrimTyConName     	      = mkPrimTc (fsLit "Int#") intPrimTyConKey  intPrimTyCon
 int32PrimTyConName	      = mkPrimTc (fsLit "Int32#") int32PrimTyConKey int32PrimTyCon
@@ -184,6 +186,7 @@ threadIdPrimTyConName  	      = mkPrimTc (fsLit "ThreadId#") threadIdPrimTyConKe
 floatX4PrimTyConName          = mkPrimTc (fsLit "FloatX4#") floatX4PrimTyConKey floatX4PrimTyCon
 doubleX2PrimTyConName         = mkPrimTc (fsLit "DoubleX2#") doubleX2PrimTyConKey doubleX2PrimTyCon
 int32X4PrimTyConName          = mkPrimTc (fsLit "Int32X4#") int32X4PrimTyConKey int32X4PrimTyCon
+int64X2PrimTyConName          = mkPrimTc (fsLit "Int64X2#") int64X2PrimTyConKey int64X2PrimTyCon
 \end{code}
 
 %************************************************************************
@@ -763,4 +766,9 @@ int32X4PrimTy :: Type
 int32X4PrimTy = mkTyConTy int32X4PrimTyCon
 int32X4PrimTyCon :: TyCon
 int32X4PrimTyCon = pcPrimTyCon0 int32X4PrimTyConName (VecRep 4 Int32ElemRep)
+
+int64X2PrimTy :: Type
+int64X2PrimTy = mkTyConTy int64X2PrimTyCon
+int64X2PrimTyCon :: TyCon
+int64X2PrimTyCon = pcPrimTyCon0 int64X2PrimTyConName (VecRep 2 Int64ElemRep)
 \end{code}
diff --git a/compiler/prelude/primops.txt.pp b/compiler/prelude/primops.txt.pp
index 03963d9b7f..5ccc162095 100644
--- a/compiler/prelude/primops.txt.pp
+++ b/compiler/prelude/primops.txt.pp
@@ -2488,6 +2488,103 @@ primop  WriteOffAddrOp_Int32AsInt32X4 "writeInt32OffAddrAsInt32X4#" GenPrimOp
    Addr# -> Int# -> Int32X4# -> State# s -> State# s
    with has_side_effects = True
         can_fail = True
+
+------------------------------------------------------------------------
+section "Int64 SIMD Vectors" 
+	{Operations on SIMD vectors of 2 64-bit signed integers.}
+------------------------------------------------------------------------
+
+primtype Int64X2#
+
+primop Int64ToInt64X2Op "int64ToInt64X2#" GenPrimOp     
+   INT64 -> Int64X2#
+
+primop Int64X2InsertOp "insertInt64X2#" GenPrimOp     
+   Int64X2# -> INT64 -> Int# -> Int64X2#
+   with can_fail = True
+
+primop Int64X2PackOp "packInt64X2#" GenPrimOp         
+   INT64 -> INT64 -> Int64X2#
+
+primop Int64X2UnpackOp "unpackInt64X2#" GenPrimOp         
+   Int64X2# -> (# INT64, INT64 #)
+
+primop Int64X2AddOp "plusInt64X2#" Dyadic            
+   Int64X2# -> Int64X2# -> Int64X2#
+   with commutable = True
+
+primop Int64X2SubOp "minusInt64X2#" Dyadic
+  Int64X2# -> Int64X2# -> Int64X2#
+
+primop Int64X2MulOp "timesInt64X2#" Dyadic    
+   Int64X2# -> Int64X2# -> Int64X2#
+   with commutable = True
+
+primop Int64X2QuotOp "quotInt64X2#" Dyadic  
+   Int64X2# -> Int64X2# -> Int64X2#
+   with can_fail = True
+   
+primop Int64X2RemOp "remInt64X2#" Dyadic  
+   Int64X2# -> Int64X2# -> Int64X2#
+   with can_fail = True
+
+primop Int64X2NegOp "negateInt64X2#" Monadic
+   Int64X2# -> Int64X2#
+
+primop IndexByteArrayOp_Int64X2 "indexInt64X2Array#" GenPrimOp
+   ByteArray# -> Int# -> Int64X2#
+   with can_fail = True
+
+primop ReadByteArrayOp_Int64X2 "readInt64X2Array#" GenPrimOp
+   MutableByteArray# s -> Int# -> State# s -> (# State# s, Int64X2# #)
+   with has_side_effects = True
+        can_fail = True
+
+primop WriteByteArrayOp_Int64X2 "writeInt64X2Array#" GenPrimOp
+   MutableByteArray# s -> Int# -> Int64X2# -> State# s -> State# s
+   with has_side_effects = True
+        can_fail = True
+
+primop IndexOffAddrOp_Int64X2 "indexInt64X2OffAddr#" GenPrimOp
+   Addr# -> Int# -> Int64X2#
+   with can_fail = True
+
+primop ReadOffAddrOp_Int64X2 "readInt64X2OffAddr#" GenPrimOp
+   Addr# -> Int# -> State# s -> (# State# s, Int64X2# #)
+   with has_side_effects = True
+
+primop  WriteOffAddrOp_Int64X2 "writeInt64X2OffAddr#" GenPrimOp
+   Addr# -> Int# -> Int64X2# -> State# s -> State# s
+   with has_side_effects = True
+        can_fail = True
+
+primop IndexByteArrayOp_Int64AsInt64X2 "indexInt64ArrayAsInt64X2#" GenPrimOp
+   ByteArray# -> Int# -> Int64X2#
+   with can_fail = True
+
+primop ReadByteArrayOp_Int64AsInt64X2 "readInt64ArrayAsInt64X2#" GenPrimOp
+   MutableByteArray# s -> Int# -> State# s -> (# State# s, Int64X2# #)
+   with has_side_effects = True
+        can_fail = True
+
+primop WriteByteArrayOp_Int64AsInt64X2 "writeInt64ArrayAsInt64X2#" GenPrimOp
+   MutableByteArray# s -> Int# -> Int64X2# -> State# s -> State# s
+   with has_side_effects = True
+        can_fail = True
+
+primop IndexOffAddrOp_Int64AsInt64X2 "indexInt64OffAddrAsInt64X2#" GenPrimOp
+   Addr# -> Int# -> Int64X2#
+   with can_fail = True
+
+primop ReadOffAddrOp_Int64AsInt64X2 "readInt64OffAddrAsInt64X2#" GenPrimOp
+   Addr# -> Int# -> State# s -> (# State# s, Int64X2# #)
+   with has_side_effects = True
+        can_fail = True
+
+primop  WriteOffAddrOp_Int64AsInt64X2 "writeInt64OffAddrAsInt64X2#" GenPrimOp
+   Addr# -> Int# -> Int64X2# -> State# s -> State# s
+   with has_side_effects = True
+        can_fail = True
    
 ------------------------------------------------------------------------
 ---                                                                  ---
diff --git a/utils/genprimopcode/Main.hs b/utils/genprimopcode/Main.hs
index 528732288b..f3d6cb8dc0 100644
--- a/utils/genprimopcode/Main.hs
+++ b/utils/genprimopcode/Main.hs
@@ -554,6 +554,7 @@ gen_wrappers (Info _ entries)
         llvmOnlyTyCon "FloatX4#"  = True
         llvmOnlyTyCon "DoubleX2#" = True
         llvmOnlyTyCon "Int32X4#"  = True
+        llvmOnlyTyCon "Int64X2#"  = True
         llvmOnlyTyCon _           = False
 
 gen_primop_list :: Info -> String
@@ -676,6 +677,7 @@ ppType (TyApp "Double#"     []) = "doublePrimTy"
 ppType (TyApp "FloatX4#"    []) = "floatX4PrimTy"
 ppType (TyApp "DoubleX2#"   []) = "doubleX2PrimTy"
 ppType (TyApp "Int32X4#"    []) = "int32X4PrimTy"
+ppType (TyApp "Int64X2#"    []) = "int64X2PrimTy"
 ppType (TyApp "ByteArray#"  []) = "byteArrayPrimTy"
 ppType (TyApp "RealWorld"   []) = "realWorldTy"
 ppType (TyApp "ThreadId#"   []) = "threadIdPrimTy"