Move entirely to array-based SIMD

See MCP#621

This tries to make as few changes as possible -- it keeps the `new` functions taking all the parameters, for example.

crates/core_arch/src/aarch64/neon/mod.rs

@@ -23,10 +23,10 @@ use stdarch_test::assert_instr;
 types! {
     /// ARM-specific 64-bit wide vector of one packed `f64`.
     #[stable(feature = "neon_intrinsics", since = "1.59.0")]
-    pub struct float64x1_t(f64); // FIXME: check this!
+    pub struct float64x1_t(1 x f64); // FIXME: check this!
     /// ARM-specific 128-bit wide vector of two packed `f64`.
     #[stable(feature = "neon_intrinsics", since = "1.59.0")]
-    pub struct float64x2_t(f64, f64);
+    pub struct float64x2_t(2 x f64);
 }
 
 /// ARM-specific type containing two `float64x1_t` vectors.

crates/core_arch/src/arm/dsp.rs

@@ -28,10 +28,10 @@ use crate::mem::transmute;
 types! {
     /// ARM-specific 32-bit wide vector of two packed `i16`.
     #[unstable(feature = "stdarch_arm_dsp", issue = "117237")]
-    pub struct int16x2_t(i16, i16);
+    pub struct int16x2_t(2 x i16);
     /// ARM-specific 32-bit wide vector of two packed `u16`.
     #[unstable(feature = "stdarch_arm_dsp", issue = "117237")]
-    pub struct uint16x2_t(u16, u16);
+    pub struct uint16x2_t(2 x u16);
 }
 
 extern "unadjusted" {