diff --git a/rust-toolchain.toml b/rust-toolchain.toml
index 7ce24e712..3e0023626 100644
--- a/rust-toolchain.toml
+++ b/rust-toolchain.toml
@@ -1,3 +1,9 @@
 [toolchain]
-channel = "nightly-2022-08-08"
+channel = "nightly-2023-05-10"
+targets = [
+    "x86_64-unknown-linux-gnu",
+    "i686-unknown-linux-gnu",
+    "aarch64-unknown-linux-gnu",
+    "armv7-unknown-linux-gnueabihf",
+]
 components = ["rustfmt"]
diff --git a/src/cdef_apply_tmpl_16.rs b/src/cdef_apply_tmpl_16.rs
index b7839718e..725ef9a00 100644
--- a/src/cdef_apply_tmpl_16.rs
+++ b/src/cdef_apply_tmpl_16.rs
@@ -929,7 +929,7 @@ pub unsafe extern "C" fn dav1d_cdef_brow_16bpc(
                 layout,
             );
         }
-        let mut lr_bak: [[[[pixel; 2]; 8]; 3]; 2] = [[[[0; 2]; 8]; 3]; 2];
+        let mut lr_bak: Align16<[[[[pixel; 2]; 8]; 3]; 2]> = Align16([[[[0; 2]; 8]; 3]; 2]);
         let mut iptrs: [*mut pixel; 3] = [ptrs[0], ptrs[1], ptrs[2]];
         edges = ::core::mem::transmute::<libc::c_uint, CdefEdgeFlags>(
             edges as libc::c_uint & !(CDEF_HAVE_LEFT as libc::c_int) as libc::c_uint,
diff --git a/src/cdef_apply_tmpl_8.rs b/src/cdef_apply_tmpl_8.rs
index eb2e077f7..9d72d7a10 100644
--- a/src/cdef_apply_tmpl_8.rs
+++ b/src/cdef_apply_tmpl_8.rs
@@ -902,7 +902,7 @@ pub unsafe extern "C" fn dav1d_cdef_brow_8bpc(
                 layout,
             );
         }
-        let mut lr_bak: [[[[pixel; 2]; 8]; 3]; 2] = [[[[0; 2]; 8]; 3]; 2];
+        let mut lr_bak: Align16<[[[[pixel; 2]; 8]; 3]; 2]> = Align16([[[[0; 2]; 8]; 3]; 2]);
         let mut iptrs: [*mut pixel; 3] = [ptrs[0], ptrs[1], ptrs[2]];
         edges = ::core::mem::transmute::<libc::c_uint, CdefEdgeFlags>(
             edges as libc::c_uint & !(CDEF_HAVE_LEFT as libc::c_int) as libc::c_uint,
diff --git a/src/cdef_tmpl_8.rs b/src/cdef_tmpl_8.rs
index 9a0507b00..ead658f1d 100644
--- a/src/cdef_tmpl_8.rs
+++ b/src/cdef_tmpl_8.rs
@@ -845,8 +845,8 @@ unsafe extern "C" fn cdef_filter_4x4_neon(
     damping: libc::c_int,
     edges: CdefEdgeFlags,
 ) {
-    let mut tmp_buf = [0; 104];
-    let mut tmp = tmp_buf.as_mut_ptr().offset(2 * 8).offset(8);
+    let mut tmp_buf = Align16([0; 104]);
+    let mut tmp = tmp_buf.0.as_mut_ptr().offset(2 * 8).offset(8);
     dav1d_cdef_padding4_8bpc_neon(tmp, dst, stride, left, top, bottom, 4, edges);
     dav1d_cdef_filter4_8bpc_neon(
         dst,
@@ -875,8 +875,8 @@ unsafe extern "C" fn cdef_filter_4x8_neon(
     damping: libc::c_int,
     edges: CdefEdgeFlags,
 ) {
-    let mut tmp_buf = [0; 104];
-    let mut tmp = tmp_buf.as_mut_ptr().offset(2 * 8).offset(8);
+    let mut tmp_buf = Align16([0; 104]);
+    let mut tmp = tmp_buf.0.as_mut_ptr().offset(2 * 8).offset(8);
     dav1d_cdef_padding4_8bpc_neon(tmp, dst, stride, left, top, bottom, 8, edges);
     dav1d_cdef_filter4_8bpc_neon(
         dst,
diff --git a/src/decode.rs b/src/decode.rs
index 65e14f14c..b22fd4445 100644
--- a/src/decode.rs
+++ b/src/decode.rs
@@ -3335,14 +3335,13 @@ unsafe fn decode_b(
     if b.skip_mode != 0 {
         b.intra = 0;
     } else if is_inter_or_switch(frame_hdr) {
-        if let Some(seg) = seg && (seg.r#ref >= 0 || seg.globalmv != 0) {
+        if let Some(seg) = seg.filter(|seg| seg.r#ref >= 0 || seg.globalmv != 0) {
             b.intra = (seg.r#ref == 0) as uint8_t;
         } else {
             let ictx = get_intra_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
-            b.intra = (!dav1d_msac_decode_bool_adapt(
-                &mut ts.msac,
-                &mut ts.cdf.m.intra[ictx.into()],
-            )) as uint8_t;
+            b.intra =
+                (!dav1d_msac_decode_bool_adapt(&mut ts.msac, &mut ts.cdf.m.intra[ictx.into()]))
+                    as uint8_t;
 
             if DEBUG_BLOCK_INFO(f, t) {
                 println!("Post-intra[{}]: r={}", b.intra, ts.msac.rng);
@@ -4266,139 +4265,64 @@ unsafe fn decode_b(
             }
         } else {
             b.c2rust_unnamed.c2rust_unnamed_0.comp_type = COMP_INTER_NONE as libc::c_int as uint8_t;
-            if let Some(seg) = seg && seg.r#ref > 0 {
-                b
-                    .c2rust_unnamed
-                    .c2rust_unnamed_0
-                    .r#ref[0 as libc::c_int
-                    as usize] = seg.r#ref as i8 - 1;
-            } else if let Some(seg) = seg && (seg.globalmv != 0 || seg.skip != 0) {
-                b
-                    .c2rust_unnamed
-                    .c2rust_unnamed_0
-                    .r#ref[0] = 0 as libc::c_int as int8_t;
+            if let Some(seg) = seg.filter(|seg| seg.r#ref > 0) {
+                b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0 as libc::c_int as usize] =
+                    seg.r#ref as i8 - 1;
+            } else if let Some(_) = seg.filter(|seg| seg.globalmv != 0 || seg.skip != 0) {
+                b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0] = 0 as libc::c_int as int8_t;
             } else {
-                let ctx1_0 = av1_get_ref_ctx(
-                    &*t.a,
-                    &t.l,
-                    by4,
-                    bx4,
-                    have_top,
-                    have_left,
-                );
+                let ctx1_0 = av1_get_ref_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
                 if dav1d_msac_decode_bool_adapt(
                     &mut ts.msac,
                     &mut ts.cdf.m.r#ref[0][ctx1_0 as usize],
-                )
-                {
-                    let ctx2_1 = av1_get_bwd_ref_ctx(
-                        &*t.a,
-                        &t.l,
-                        by4,
-                        bx4,
-                        have_top,
-                        have_left,
-                    );
+                ) {
+                    let ctx2_1 = av1_get_bwd_ref_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
                     if dav1d_msac_decode_bool_adapt(
                         &mut ts.msac,
                         &mut ts.cdf.m.r#ref[1][ctx2_1 as usize],
-                    )
-                    {
-                        b
-                            .c2rust_unnamed
-                            .c2rust_unnamed_0
-                            .r#ref[0 as libc::c_int
-                            as usize] = 6 as libc::c_int as int8_t;
+                    ) {
+                        b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0 as libc::c_int as usize] =
+                            6 as libc::c_int as int8_t;
                     } else {
-                        let ctx3_0 = av1_get_bwd_ref_1_ctx(
-                            &*t.a,
-                            &t.l,
-                            by4,
-                            bx4,
-                            have_top,
-                            have_left,
-                        );
-                        b
-                            .c2rust_unnamed
-                            .c2rust_unnamed_0
-                            .r#ref[0 as libc::c_int
-                            as usize] = (4 as libc::c_int as libc::c_uint)
-                            .wrapping_add(
+                        let ctx3_0 =
+                            av1_get_bwd_ref_1_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
+                        b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0 as libc::c_int as usize] =
+                            (4 as libc::c_int as libc::c_uint).wrapping_add(
                                 dav1d_msac_decode_bool_adapt(
                                     &mut ts.msac,
-                                    &mut ts
-                                        .cdf
-                                        .m
-                                        .r#ref[5][ctx3_0 as usize],
+                                    &mut ts.cdf.m.r#ref[5][ctx3_0 as usize],
                                 ) as libc::c_uint,
                             ) as int8_t;
                     }
                 } else {
-                    let ctx2_2 = av1_get_fwd_ref_ctx(
-                        &*t.a,
-                        &t.l,
-                        by4,
-                        bx4,
-                        have_top,
-                        have_left,
-                    );
+                    let ctx2_2 = av1_get_fwd_ref_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
                     if dav1d_msac_decode_bool_adapt(
                         &mut ts.msac,
                         &mut ts.cdf.m.r#ref[2][ctx2_2 as usize],
-                    )
-                    {
-                        let ctx3_1 = av1_get_fwd_ref_2_ctx(
-                            &*t.a,
-                            &t.l,
-                            by4,
-                            bx4,
-                            have_top,
-                            have_left,
-                        );
-                        b
-                            .c2rust_unnamed
-                            .c2rust_unnamed_0
-                            .r#ref[0 as libc::c_int
-                            as usize] = (2 as libc::c_int as libc::c_uint)
-                            .wrapping_add(
+                    ) {
+                        let ctx3_1 =
+                            av1_get_fwd_ref_2_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
+                        b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0 as libc::c_int as usize] =
+                            (2 as libc::c_int as libc::c_uint).wrapping_add(
                                 dav1d_msac_decode_bool_adapt(
                                     &mut ts.msac,
-                                    &mut ts
-                                        .cdf
-                                        .m
-                                        .r#ref[4][ctx3_1 as usize],
+                                    &mut ts.cdf.m.r#ref[4][ctx3_1 as usize],
                                 ) as libc::c_uint,
                             ) as int8_t;
                     } else {
-                        let ctx3_2 = av1_get_fwd_ref_1_ctx(
-                            &*t.a,
-                            &t.l,
-                            by4,
-                            bx4,
-                            have_top,
-                            have_left,
-                        );
-                        b
-                            .c2rust_unnamed
-                            .c2rust_unnamed_0
-                            .r#ref[0 as libc::c_int
-                            as usize] = dav1d_msac_decode_bool_adapt(
-                            &mut ts.msac,
-                            &mut ts
-                                .cdf
-                                .m
-                                .r#ref[3][ctx3_2 as usize],
-                        ) as int8_t;
+                        let ctx3_2 =
+                            av1_get_fwd_ref_1_ctx(&*t.a, &t.l, by4, bx4, have_top, have_left);
+                        b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0 as libc::c_int as usize] =
+                            dav1d_msac_decode_bool_adapt(
+                                &mut ts.msac,
+                                &mut ts.cdf.m.r#ref[3][ctx3_2 as usize],
+                            ) as int8_t;
                     }
                 }
-                if DEBUG_BLOCK_INFO(f, t)
-                {
+                if DEBUG_BLOCK_INFO(f, t) {
                     printf(
                         b"Post-ref[%d]: r=%d\n\0" as *const u8 as *const libc::c_char,
-                        b
-                            .c2rust_unnamed
-                            .c2rust_unnamed_0
-                            .r#ref[0] as libc::c_int,
+                        b.c2rust_unnamed.c2rust_unnamed_0.r#ref[0] as libc::c_int,
                         ts.msac.rng,
                     );
                 }
diff --git a/src/env.rs b/src/env.rs
index 5dc7b164d..57778faa1 100644
--- a/src/env.rs
+++ b/src/env.rs
@@ -578,9 +578,10 @@ pub fn av1_get_uni_p1_ctx(
         if let Some(cnt) = cnt.get_mut((a.r#ref[0][xb4 as usize] - 1) as usize) {
             *cnt += 1;
         }
-        if a.comp_type[xb4 as usize] != 0
-                && let Some(cnt) = cnt.get_mut((a.r#ref[1][xb4 as usize] - 1) as usize) {
-            *cnt += 1;
+        if a.comp_type[xb4 as usize] != 0 {
+            if let Some(cnt) = cnt.get_mut((a.r#ref[1][xb4 as usize] - 1) as usize) {
+                *cnt += 1;
+            }
         }
     }
 
@@ -588,9 +589,10 @@ pub fn av1_get_uni_p1_ctx(
         if let Some(cnt) = cnt.get_mut((l.r#ref[0][yb4 as usize] - 1) as usize) {
             *cnt += 1;
         }
-        if l.comp_type[yb4 as usize] != 0
-                && let Some(cnt) = cnt.get_mut((l.r#ref[1][yb4 as usize] - 1) as usize) {
-            *cnt += 1;
+        if l.comp_type[yb4 as usize] != 0 {
+            if let Some(cnt) = cnt.get_mut((l.r#ref[1][yb4 as usize] - 1) as usize) {
+                *cnt += 1;
+            }
         }
     }
 
diff --git a/src/fg_apply_tmpl_16.rs b/src/fg_apply_tmpl_16.rs
index aaba86941..60278a5ed 100644
--- a/src/fg_apply_tmpl_16.rs
+++ b/src/fg_apply_tmpl_16.rs
@@ -1,6 +1,6 @@
 use crate::include::stddef::*;
 use crate::include::stdint::*;
-use crate::src::align::{Align16, Align64};
+use crate::src::align::{Align1, Align16};
 use ::libc;
 extern "C" {
     pub type Dav1dRef;
@@ -392,7 +392,7 @@ pub unsafe extern "C" fn dav1d_apply_grain_16bpc(
     in_0: *const Dav1dPicture,
 ) {
     let mut grain_lut = Align16([[[0; 82]; 74]; 3]);
-    let mut scaling = Align64([[0; 4096]; 3]);
+    let mut scaling = Align1([[0; 4096]; 3]);
     let rows = (*out).p.h + 31 >> 5;
     dav1d_prep_grain_16bpc(
         dsp,
diff --git a/src/fg_apply_tmpl_8.rs b/src/fg_apply_tmpl_8.rs
index 4deec89c1..c2619b763 100644
--- a/src/fg_apply_tmpl_8.rs
+++ b/src/fg_apply_tmpl_8.rs
@@ -1,7 +1,8 @@
 use crate::include::stddef::*;
 use crate::include::stdint::*;
-use crate::src::align::{Align16, Align64};
+use crate::src::align::Align16;
 use ::libc;
+use cfg_if::cfg_if;
 extern "C" {
     pub type Dav1dRef;
     fn memcpy(_: *mut libc::c_void, _: *const libc::c_void, _: libc::c_ulong) -> *mut libc::c_void;
@@ -348,7 +349,17 @@ pub unsafe extern "C" fn dav1d_apply_grain_8bpc(
     in_0: *const Dav1dPicture,
 ) {
     let mut grain_lut = Align16([[[0; 82]; 74]; 3]);
-    let mut scaling = Align64([[0; 256]; 3]);
+    cfg_if! {
+        if #[cfg(target_arch = "x86_64")] {
+            use crate::src::align::Align64;
+
+            let mut scaling = Align64([[0; 256]; 3]);
+        } else {
+            use crate::src::align::Align1;
+
+            let mut scaling = Align1([[0; 256]; 3]);
+        }
+    }
     let rows = (*out).p.h + 31 >> 5;
     dav1d_prep_grain_8bpc(
         dsp,
diff --git a/src/lf_mask.rs b/src/lf_mask.rs
index ea6826719..829e75e4d 100644
--- a/src/lf_mask.rs
+++ b/src/lf_mask.rs
@@ -60,6 +60,9 @@ use crate::include::common::intops::iclip;
 use crate::include::common::intops::imax;
 use crate::include::common::intops::imin;
 use crate::src::tables::TxfmInfo;
+
+use crate::src::align::Align16;
+
 unsafe extern "C" fn decomp_tx(
     txa: *mut [[[uint8_t; 32]; 32]; 2],
     from: RectTxfmSize,
@@ -324,7 +327,7 @@ unsafe extern "C" fn mask_edges_inter(
         &*dav1d_txfm_dimensions.as_ptr().offset(max_tx as isize) as *const TxfmInfo;
     let mut y = 0;
     let mut x = 0;
-    let mut txa: [[[[uint8_t; 32]; 32]; 2]; 2] = [[[[0; 32]; 32]; 2]; 2];
+    let mut txa: Align16<[[[[uint8_t; 32]; 32]; 2]; 2]> = Align16([[[[0; 32]; 32]; 2]; 2]);
     let mut y_off = 0;
     let mut y_0 = 0;
     while y_0 < h4 {
@@ -332,7 +335,7 @@ unsafe extern "C" fn mask_edges_inter(
         let mut x_0 = 0;
         while x_0 < w4 {
             decomp_tx(
-                &mut *(*(*(*txa.as_mut_ptr().offset(0)).as_mut_ptr().offset(0))
+                &mut *(*(*(*txa.0.as_mut_ptr().offset(0)).as_mut_ptr().offset(0))
                     .as_mut_ptr()
                     .offset(y_0 as isize))
                 .as_mut_ptr()
diff --git a/src/lr_apply_tmpl_16.rs b/src/lr_apply_tmpl_16.rs
index bef395e21..073e3b029 100644
--- a/src/lr_apply_tmpl_16.rs
+++ b/src/lr_apply_tmpl_16.rs
@@ -869,7 +869,7 @@ unsafe extern "C" fn lr_sbrow(
     let max_unit_size = unit_size + half_unit_size;
     let row_y = y + (8 >> ss_ver) * (y != 0) as libc::c_int;
     let shift_hor = 7 - ss_hor;
-    let mut pre_lr_border: [[[pixel; 4]; 136]; 2] = [[[0; 4]; 136]; 2];
+    let mut pre_lr_border: Align16<[[[pixel; 4]; 136]; 2]> = Align16([[[0; 4]; 136]; 2]);
     let mut lr: [*const Av1RestorationUnit; 2] = [0 as *const Av1RestorationUnit; 2];
     let mut edges: LrEdgeFlags = ((if y > 0 {
         LR_HAVE_TOP as libc::c_int
diff --git a/src/lr_apply_tmpl_8.rs b/src/lr_apply_tmpl_8.rs
index f55055f01..56d53077d 100644
--- a/src/lr_apply_tmpl_8.rs
+++ b/src/lr_apply_tmpl_8.rs
@@ -840,7 +840,7 @@ unsafe extern "C" fn lr_sbrow(
     let max_unit_size = unit_size + half_unit_size;
     let row_y = y + (8 >> ss_ver) * (y != 0) as libc::c_int;
     let shift_hor = 7 - ss_hor;
-    let mut pre_lr_border: [[[pixel; 4]; 136]; 2] = [[[0; 4]; 136]; 2];
+    let mut pre_lr_border: Align16<[[[pixel; 4]; 136]; 2]> = Align16([[[0; 4]; 136]; 2]);
     let mut lr: [*const Av1RestorationUnit; 2] = [0 as *const Av1RestorationUnit; 2];
     let mut edges: LrEdgeFlags = ((if y > 0 {
         LR_HAVE_TOP as libc::c_int