cg_ssa: transmute between scalable vectors

Like regular SIMD vectors, we can support casting between scalable
vectors of integral or floating-point types without needing a temporary.

compiler/rustc_codegen_ssa/src/mir/rvalue.rs

@@ -302,6 +302,14 @@ fn vector_can_bitcast(x: abi::Scalar) -> bool {
                 let to_backend_ty = bx.cx().immediate_backend_type(cast);
                 OperandValue::Immediate(bx.bitcast(imm, to_backend_ty))
             }
+            (
+                OperandValue::Immediate(imm),
+                abi::BackendRepr::SimdScalableVector { element: from_scalar, .. },
+                abi::BackendRepr::SimdScalableVector { element: to_scalar, .. },
+            ) if vector_can_bitcast(from_scalar) && vector_can_bitcast(to_scalar) => {
+                let to_backend_ty = bx.cx().immediate_backend_type(cast);
+                OperandValue::Immediate(bx.bitcast(imm, to_backend_ty))
+            }
             (
                 OperandValue::Pair(imm_a, imm_b),
                 abi::BackendRepr::ScalarPair(in_a, in_b),

tests/ui/scalable-vectors/transmute.rs

@@ -0,0 +1,39 @@
+//@ build-pass
+//@ compile-flags: -Copt-level=3
+//@ only-aarch64
+#![crate_type = "lib"]
+#![allow(incomplete_features, internal_features, dead_code, improper_ctypes)]
+#![allow(nonstandard_style, private_interfaces)]
+#![feature(abi_unadjusted, link_llvm_intrinsics, rustc_attrs)]
+
+// Tests that use of transmute between `svuint8x2_t` and `svint8x2_t` builds with optimisations
+// without any failures from LLVM.
+
+use std::mem::transmute;
+
+#[rustc_scalable_vector(16)]
+struct svbool_t(bool);
+
+#[rustc_scalable_vector(16)]
+struct svuint8_t(u8);
+
+#[rustc_scalable_vector]
+struct svuint8x2_t(svuint8_t, svuint8_t);
+
+#[rustc_scalable_vector(16)]
+struct svint8_t(i8);
+
+#[rustc_scalable_vector]
+struct svint8x2_t(svint8_t, svint8_t);
+
+#[target_feature(enable = "sve")]
+pub unsafe fn svld2_u8(pg: svbool_t, base: *const i8) -> svuint8x2_t {
+    unsafe extern "unadjusted" {
+        #[cfg_attr(
+            target_arch = "aarch64",
+            link_name = "llvm.aarch64.sve.ld2.sret.nxv16i8"
+        )]
+        fn _svld2_s8(pg: svbool_t, base: *const i8) -> svint8x2_t;
+    }
+    transmute(_svld2_s8(pg, base))
+}