debuginfo: no spill <vscale x N x i1> for N!=16

LLVM doesn't handle stores on `<vscale x N x i1>` for `N != 16`, a type
used internally in SVE intrinsics. Spilling to the stack to create
debuginfo will cause errors during instruction selection. These types
that are an internal implementation detail to the intrinsic, so users
should never see them types and won't need any debuginfo.

Co-authored-by: Jamie Cunliffe <Jamie.Cunliffe@arm.com>

Author: davidtwco

compiler/rustc_codegen_ssa/src/mir/debuginfo.rs

@@ -361,6 +361,49 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     return;
                 }
 
+                // Don't spill `<vscale x N x i1>` for `N != 16`:
+                //
+                // SVE predicates are only one bit for each byte in an SVE vector (which makes
+                // sense, the predicate only needs to keep track of whether a lane is
+                // enabled/disabled). i.e. a `<vscale x 16 x i8>` vector has a `<vscale x 16 x i1>`
+                // predicate type. `<vscale x 16 x i1>` corresponds to two bytes of storage,
+                // multiplied by the `vscale`, with one bit for each of the sixteen lanes.
+                //
+                // For a vector with fewer elements, such as `svint32_t`/`<vscale x 4 x i32>`,
+                // while only a `<vscale x 4 x i1>` predicate type would be strictly necessary,
+                // relevant intrinsics still take a `svbool_t`/`<vscale x 16 x i1>` - this is
+                // because a `<vscale x 4 x i1>` is only half of a byte (for `vscale=1`), and with
+                // memory being byte-addressable, it's unclear how to store that.
+                //
+                // Due to this, LLVM ultimately decided not to support stores of `<vscale x N x i1>`
+                // for `N != 16`. As for `vscale=1` and `N` fewer than sixteen, partial bytes would
+                // need to be stored (except for `N=8`, but that also isn't supported). `N` can
+                // never be greater than sixteen as that ends up larger than the 128-bit increment
+                // size.
+                //
+                // Internally, with an intrinsic operating on a `svint32_t`/`<vscale x 4 x i32>`
+                // (for example), the intrinsic takes the `svbool_t`/`<vscale x 16 x i1>` predicate
+                // and casts it to a `svbool4_t`/`<vscale x 4 x i1>`. Therefore, it's important that
+                // the `<vscale x 4 x i32>` never spills because that'll cause errors during
+                // instruction selection. Spilling to the stack to create debuginfo for these
+                // intermediate values must be avoided and won't degrade the debugging experience
+                // anyway.
+                if operand.layout.ty.is_scalable_simd()
+                    && bx.sess().target.arch == "aarch64"
+                    && let ty::Adt(adt, args) = &operand.layout.ty.kind()
+                    && let Some(marker_type_field) =
+                        adt.non_enum_variant().fields.get(FieldIdx::from_u32(0))
+                {
+                    let marker_type = marker_type_field.ty(bx.tcx(), args);
+                    // i.e. `<vscale x N x i1>` when `N != 16`
+                    if let ty::Slice(element_ty) = marker_type.kind()
+                        && element_ty.is_bool()
+                        && adt.repr().scalable != Some(16)
+                    {
+                        return;
+                    }
+                }
+
                 Self::spill_operand_to_stack(*operand, name, bx)
             }