Fix missing stack aligns

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,

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,

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,

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,

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,

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,
@@ -324,15 +327,15 @@ 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 {
         let mut x_off = 0;
         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()

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

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