Merge pull request #3272 from dheaton-arm/implement-iaddpairwise

Implement `IaddPairwise` for the interpreter

cranelift/filetests/filetests/runtests/simd-iaddpairwise.clif

@@ -0,0 +1,25 @@
+test interpret
+test run
+target aarch64
+
+function %iaddp_i8x16(i8x16, i8x16) -> i8x16 {
+block0(v0: i8x16, v1: i8x16):
+    v2 = iadd_pairwise v0, v1
+    return v2
+}
+; run: %iaddp_i8x16([1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16], [17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32]) == [3 7 11 15 19 23 27 31 35 39 43 47 51 55 59 63]
+
+function %iaddp_i16x8(i16x8, i16x8) -> i16x8 {
+block0(v0: i16x8, v1: i16x8):
+    v2 = iadd_pairwise v0, v1
+    return v2
+}
+; run: %iaddp_i16x8([1 2 3 4 5 6 7 8], [100 99 98 97 96 95 94 93]) == [3 7 11 15 199 195 191 187]
+
+function %iaddp_i32x4(i32x4, i32x4) -> i32x4 {
+block0(v0: i32x4, v1: i32x4):
+    v2 = iadd_pairwise v0, v1
+    return v2
+}
+; run: %iaddp_i32x4([1 2 3 4], [5 6 7 8]) == [3 7 11 15]
+; run: %iaddp_i32x4([4294967290 5 4294967290 5], [100 100 100 100]) == [4294967295 4294967295 200 200]

cranelift/interpreter/src/step.rs

@@ -745,7 +745,7 @@ where
         Opcode::Fence => unimplemented!("Fence"),
         Opcode::WideningPairwiseDotProductS => unimplemented!("WideningPairwiseDotProductS"),
         Opcode::SqmulRoundSat => unimplemented!("SqmulRoundSat"),
-        Opcode::IaddPairwise => unimplemented!("IaddPairwise"),
+        Opcode::IaddPairwise => assign(binary_pairwise(arg(0)?, arg(1)?, ctrl_ty, Value::add)?),
 
         // TODO: these instructions should be removed once the new backend makes these obsolete
         // (see https://github.com/bytecodealliance/wasmtime/issues/1936); additionally, the
@@ -998,3 +998,24 @@ where
     }
     vectorizelanes(&result, vector_type)
 }
+
+/// Performs the supplied pairwise arithmetic `op` on two SIMD vectors, where
+/// pairs are formed from adjacent vector elements and the vectors are
+/// concatenated at the end.
+fn binary_pairwise<V, F>(x: V, y: V, vector_type: types::Type, op: F) -> ValueResult<V>
+where
+    V: Value,
+    F: Fn(V, V) -> ValueResult<V>,
+{
+    let arg0 = extractlanes(&x, vector_type.lane_type())?;
+    let arg1 = extractlanes(&y, vector_type.lane_type())?;
+    let mut result = SimdVec::new();
+    for pair in arg0.chunks(2).chain(arg1.chunks(2)) {
+        let lhs: V = Value::int(pair[0], vector_type.lane_type())?;
+        let rhs: V = Value::int(pair[1], vector_type.lane_type())?;
+        let sum = op(lhs, rhs)?;
+        let sum = sum.into_int()?;
+        result.push(sum);
+    }
+    vectorizelanes(&result, vector_type)
+}