[Hexagon] Bitcast v4i16 to v8i8, unify no-op casts between scalar and HVX

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@372616 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/Hexagon/HexagonPatterns.td b/lib/Target/Hexagon/HexagonPatterns.td
index c54be2f16f1d6daebbcd691c0e1c96c24ef107e1..485e658e1c8469de10a609eb8b7eadc8f0dafab2 100644 (file)
--- a/lib/Target/Hexagon/HexagonPatterns.td
+++ b/lib/Target/Hexagon/HexagonPatterns.td
@@ -344,6 +344,13 @@ multiclass MinMax_pats<InstHexagon PickT, InstHexagon PickS,
            (PickS CmpPred:$Vs, CmpPred:$Vt)>;
 }
 
+// Bitcasts between same-size vector types are no-ops, except for the
+// actual type change.
+multiclass NopCast_pat<ValueType Ty1, ValueType Ty2, RegisterClass RC> {
+  def: Pat<(Ty1 (bitconvert (Ty2 RC:$Val))), (Ty1 RC:$Val)>;
+  def: Pat<(Ty2 (bitconvert (Ty1 RC:$Val))), (Ty2 RC:$Val)>;
+}
+
 
 // Frags for commonly used SDNodes.
 def Add: pf2<add>;    def And: pf2<and>;    def Sra: pf2<sra>;
@@ -427,17 +434,18 @@ def: Pat<(f32 (bitconvert I32:$v)), (F32:$v)>;
 def: Pat<(i64 (bitconvert F64:$v)), (I64:$v)>;
 def: Pat<(f64 (bitconvert I64:$v)), (F64:$v)>;
 
-multiclass Cast_pat<ValueType Ta, ValueType Tb, RegisterClass RC> {
-  def: Pat<(Tb (bitconvert (Ta RC:$Rs))), (Tb RC:$Rs)>;
-  def: Pat<(Ta (bitconvert (Tb RC:$Rs))), (Ta RC:$Rs)>;
-}
-
-// Bit convert vector types to integers.
-defm: Cast_pat<v4i8,  i32, IntRegs>;
-defm: Cast_pat<v2i16, i32, IntRegs>;
-defm: Cast_pat<v8i8,  i64, DoubleRegs>;
-defm: Cast_pat<v4i16, i64, DoubleRegs>;
-defm: Cast_pat<v2i32, i64, DoubleRegs>;
+// Bit convert 32- and 64-bit types.
+// All of these are bitcastable to one another: i32, v2i16, v4i8.
+defm: NopCast_pat<i32,   v2i16, IntRegs>;
+defm: NopCast_pat<i32,    v4i8, IntRegs>;
+defm: NopCast_pat<v2i16,  v4i8, IntRegs>;
+// All of these are bitcastable to one another: i64, v2i32, v4i16, v8i8.
+defm: NopCast_pat<i64,   v2i32, DoubleRegs>;
+defm: NopCast_pat<i64,   v4i16, DoubleRegs>;
+defm: NopCast_pat<i64,    v8i8, DoubleRegs>;
+defm: NopCast_pat<v2i32, v4i16, DoubleRegs>;
+defm: NopCast_pat<v2i32,  v8i8, DoubleRegs>;
+defm: NopCast_pat<v4i16,  v8i8, DoubleRegs>;
 
 
 // --(3) Extend/truncate -------------------------------------------------

diff --git a/lib/Target/Hexagon/HexagonPatternsHVX.td b/lib/Target/Hexagon/HexagonPatternsHVX.td
index 32e2260de07c066fa700a57b31efea6a827ceba0..078a7135c55be40a6eb4a18b9ed980c9f4c677ec 100644 (file)
--- a/lib/Target/Hexagon/HexagonPatternsHVX.td
+++ b/lib/Target/Hexagon/HexagonPatternsHVX.td
@@ -160,23 +160,14 @@ let Predicates = [UseHVX] in {
 
 // Bitcasts between same-size vector types are no-ops, except for the
 // actual type change.
-class Bitcast<ValueType ResTy, ValueType InpTy, RegisterClass RC>
-  : Pat<(ResTy (bitconvert (InpTy RC:$Val))), (ResTy RC:$Val)>;
-
 let Predicates = [UseHVX] in {
-  def: Bitcast<VecI8,   VecI16,  HvxVR>;
-  def: Bitcast<VecI8,   VecI32,  HvxVR>;
-  def: Bitcast<VecI16,  VecI8,   HvxVR>;
-  def: Bitcast<VecI16,  VecI32,  HvxVR>;
-  def: Bitcast<VecI32,  VecI8,   HvxVR>;
-  def: Bitcast<VecI32,  VecI16,  HvxVR>;
-
-  def: Bitcast<VecPI8,  VecPI16, HvxWR>;
-  def: Bitcast<VecPI8,  VecPI32, HvxWR>;
-  def: Bitcast<VecPI16, VecPI8,  HvxWR>;
-  def: Bitcast<VecPI16, VecPI32, HvxWR>;
-  def: Bitcast<VecPI32, VecPI8,  HvxWR>;
-  def: Bitcast<VecPI32, VecPI16, HvxWR>;
+  defm: NopCast_pat<VecI8,   VecI16,  HvxVR>;
+  defm: NopCast_pat<VecI8,   VecI32,  HvxVR>;
+  defm: NopCast_pat<VecI16,  VecI32,  HvxVR>;
+
+  defm: NopCast_pat<VecPI8,  VecPI16, HvxWR>;
+  defm: NopCast_pat<VecPI8,  VecPI32, HvxWR>;
+  defm: NopCast_pat<VecPI16, VecPI32, HvxWR>;
 }
 
 let Predicates = [UseHVX] in {

diff --git a/test/CodeGen/Hexagon/isel-bitcast-v8i8-v4i16.ll b/test/CodeGen/Hexagon/isel-bitcast-v8i8-v4i16.ll
new file mode 100644 (file)
index 0000000..f45b1a7
--- /dev/null
+++ b/test/CodeGen/Hexagon/isel-bitcast-v8i8-v4i16.ll
@@ -0,0 +1,13 @@
+; RUN: llc -march=hexagon < %s | FileCheck %s
+
+; Check that this doesn't fail to select instructions.
+; CHECK: vsplath
+
+define <8 x i8> @fred(i16 %a0) #0 {
+  %t0 = insertelement <4 x i16> undef, i16 %a0, i32 0
+  %t1 = shufflevector <4 x i16> %t0, <4 x i16> undef, <4 x i32> zeroinitializer
+  %t2 = bitcast <4 x i16> %t1 to <8 x i8>
+  ret <8 x i8> %t2
+}
+
+attributes #0 = { readnone nounwind "target-cpu"="hexagonv62" }