From ff542e86f3d0045240dac379ed50d6dccedf7987 Mon Sep 17 00:00:00 2001
From: Khyber Sen <kkysen@gmail.com>
Date: Thu, 9 Nov 2023 14:35:43 -0800
Subject: [PATCH] `fn fguv_32x32xn_*`: Make `{uv,is_uv,uv_pl}` args `bool`s,
 casting to `c_int`/`usize` for FFI/indexing.

By making the arg a `bool` in `fn fguv_32x32xn_rust` and doing the `as usize` cast inside the `fn`[1],
(4[2]) bounds checks of `uv` into length-2 arrays are eliminated, fixing a perf regression.
This was checked on `aarch64-unknown-linux-gnu` asm.

[1] Doing the `bool` to `usize` cast inside of the `add_noise_uv` closure didn't make a difference in the number of bounds checks.
[2] 27 => 23 `bl core::panicking::panic_bounds_check`s
---
 src/fg_apply.rs  |  4 ++--
 src/filmgrain.rs | 10 +++++-----
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/src/fg_apply.rs b/src/fg_apply.rs
index ae10a1565..41844d55f 100644
--- a/src/fg_apply.rs
+++ b/src/fg_apply.rs
@@ -238,7 +238,7 @@ pub(crate) unsafe fn rav1d_apply_grain_row<BD: BitDepth>(
                 row,
                 luma_src,
                 r#in.stride[0],
-                pl,
+                pl != 0,
                 is_id,
                 bd,
             );
@@ -258,7 +258,7 @@ pub(crate) unsafe fn rav1d_apply_grain_row<BD: BitDepth>(
                     row,
                     luma_src,
                     r#in.stride[0],
-                    pl,
+                    pl != 0,
                     is_id,
                     bd,
                 );
diff --git a/src/filmgrain.rs b/src/filmgrain.rs
index 4603c305f..fc7648324 100644
--- a/src/filmgrain.rs
+++ b/src/filmgrain.rs
@@ -153,7 +153,7 @@ impl FnFGUV32x32xN {
         row_num: usize,
         luma_row: *const BD::Pixel,
         luma_stride: ptrdiff_t,
-        uv_pl: usize,
+        is_uv: bool,
         is_id: bool,
         bd: BD,
     ) {
@@ -165,7 +165,7 @@ impl FnFGUV32x32xN {
         let bh = bh as c_int;
         let row_num = row_num as c_int;
         let luma_row = luma_row.cast();
-        let uv_pl = uv_pl as c_int;
+        let uv_pl = is_uv as c_int;
         let is_id = is_id as c_int;
         let bd = bd.into_c();
         (self.get())(
@@ -706,13 +706,13 @@ unsafe fn fguv_32x32xn_rust<BD: BitDepth>(
     row_num: usize,
     luma_row: *const BD::Pixel,
     luma_stride: ptrdiff_t,
-    uv: usize,
+    is_uv: bool,
     is_id: bool,
     is_sx: bool,
     is_sy: bool,
     bd: BD,
 ) {
-    let [sx, sy] = [is_sx, is_sy].map(|it| it as usize);
+    let [uv, sx, sy] = [is_uv, is_sx, is_sy].map(|it| it as usize);
 
     let rows = 1 + (data.overlap_flag && row_num > 0) as usize;
     let bitdepth_min_8 = bd.bitdepth() - 8;
@@ -894,7 +894,7 @@ unsafe extern "C" fn fguv_32x32xn_c_erased<
         row_num,
         luma_row,
         luma_stride,
-        uv_pl,
+        uv_pl != 0,
         is_id,
         IS_SX,
         IS_SY,