From b31d11b379b717d4d9dfb963ca44396b731da324 Mon Sep 17 00:00:00 2001
From: Paolo Teti <paolo.teti@gmail.com>
Date: Sat, 23 Feb 2019 20:09:17 +0100
Subject: [PATCH 1/2] ACLE/SIMD32: add `ssub8` and `usub8`

- add `ssub8` and `usub8`
- bump instruction limit to 29
---
 crates/core_arch/src/acle/simd32.rs | 67 +++++++++++++++++++++++++++--
 crates/stdsimd-test/src/lib.rs      |  2 +-
 2 files changed, 65 insertions(+), 4 deletions(-)

diff --git a/crates/core_arch/src/acle/simd32.rs b/crates/core_arch/src/acle/simd32.rs
index 59a1f0a47d..e5d00e1451 100644
--- a/crates/core_arch/src/acle/simd32.rs
+++ b/crates/core_arch/src/acle/simd32.rs
@@ -16,13 +16,13 @@
 //! - [x] __sadd8
 //! - [x] __shadd8
 //! - [x] __shsub8
-//! - [ ] __ssub8
+//! - [x] __ssub8
 //! - [ ] __uadd8
 //! - [ ] __uhadd8
 //! - [ ] __uhsub8
 //! - [ ] __uqadd8
 //! - [ ] __uqsub8
-//! - [ ] __usub8
+//! - [x] __usub8
 //! - [x] __usad8
 //! - [x] __usada8
 //! - [x] __qadd16
@@ -130,6 +130,12 @@ extern "C" {
     #[link_name = "llvm.arm.shsub8"]
     fn arm_shsub8(a: i32, b: i32) -> i32;
 
+    #[link_name = "llvm.arm.ssub8"]
+    fn arm_ssub8(a: i32, b: i32) -> i32;
+
+    #[link_name = "llvm.arm.usub8"]
+    fn arm_usub8(a: i32, b: i32) -> i32;
+
     #[link_name = "llvm.arm.shsub16"]
     fn arm_shsub16(a: i32, b: i32) -> i32;
 
@@ -337,6 +343,39 @@ pub unsafe fn __shsub8(a: int8x4_t, b: int8x4_t) -> int8x4_t {
     dsp_call!(arm_shsub8, a, b)
 }
 
+/// Inserts a `USUB8` instruction.
+///
+/// Returns the 8-bit unsigned equivalent of
+///
+/// res\[0\] = a\[0\] - a\[0\]
+/// res\[1\] = a\[1\] - a\[1\]
+/// res\[2\] = a\[2\] - a\[2\]
+/// res\[3\] = a\[3\] - a\[3\]
+///
+/// where [0] is the lower 8 bits and [3] is the upper 8 bits.
+#[inline]
+#[cfg_attr(test, assert_instr(usub8))]
+pub unsafe fn __usub8(a: uint8x4_t, b: uint8x4_t) -> uint8x4_t {
+    dsp_call!(arm_usub8, a, b)
+}
+
+/// Inserts a `SSUB8` instruction.
+///
+/// Returns the 8-bit signed equivalent of
+///
+/// res\[0\] = a\[0\] - a\[0\]
+/// res\[1\] = a\[1\] - a\[1\]
+/// res\[2\] = a\[2\] - a\[2\]
+/// res\[3\] = a\[3\] - a\[3\]
+///
+/// where [0] is the lower 8 bits and [3] is the upper 8 bits.
+/// The GE bits of the APSR are set.
+#[inline]
+#[cfg_attr(test, assert_instr(ssub8))]
+pub unsafe fn __ssub8(a: int8x4_t, b: int8x4_t) -> int8x4_t {
+    dsp_call!(arm_ssub8, a, b)
+}
+
 /// Signed halving parallel halfword-wise subtraction.
 ///
 /// Returns the 16-bit signed equivalent of
@@ -427,7 +466,7 @@ pub unsafe fn __usada8(a: int8x4_t, b: int8x4_t, c: u32) -> u32 {
 
 #[cfg(test)]
 mod tests {
-    use crate::core_arch::simd::{i16x2, i8x4};
+    use crate::core_arch::simd::{i16x2, i8x4, u8x4};
     use std::mem::transmute;
     use stdsimd_test::simd_test;
 
@@ -596,6 +635,28 @@ mod tests {
         }
     }
 
+    #[test]
+    fn ssub8() {
+        unsafe {
+            let a = i8x4::new(1, 2, 3, 4);
+            let b = i8x4::new(5, 4, 3, 2);
+            let c = i8x4::new(-4, -2, 0, 2);
+            let r: i8x4 = dsp_call!(super::__ssub8, a, b);
+            assert_eq!(r, c);
+        }
+    }
+
+    #[test]
+    fn usub8() {
+        unsafe {
+            let a = u8x4::new(1, 2, 3, 4);
+            let b = u8x4::new(5, 4, 3, 2);
+            let c = u8x4::new(252, 254, 0, 2);
+            let r: u8x4 = dsp_call!(super::__usub8, a, b);
+            assert_eq!(r, c);
+        }
+    }
+
     #[test]
     fn shsub16() {
         unsafe {
diff --git a/crates/stdsimd-test/src/lib.rs b/crates/stdsimd-test/src/lib.rs
index 05d9199933..af1a9a2806 100644
--- a/crates/stdsimd-test/src/lib.rs
+++ b/crates/stdsimd-test/src/lib.rs
@@ -153,7 +153,7 @@ pub fn assert(fnptr: usize, fnname: &str, expected: &str) {
 
                 // core_arch/src/acle/simd32
                 "usad8" => 27,
-                "qadd8" | "qsub8" | "sadd8" | "sel" | "shadd8" | "shsub8" => 29,
+                "qadd8" | "qsub8" | "sadd8" | "sel" | "shadd8" | "shsub8" | "usub8" | "ssub8" => 29,
 
                 // Original limit was 20 instructions, but ARM DSP Intrinsics
                 // are exactly 20 instructions long. So, bump the limit to 22

From 53115977197f526225f6cf35ebcabc234e095802 Mon Sep 17 00:00:00 2001
From: Paolo Teti <paolo.teti@gmail.com>
Date: Mon, 25 Feb 2019 19:26:32 +0100
Subject: [PATCH 2/2] The GE bits of the APSR are set also by USUB8

---
 crates/core_arch/src/acle/simd32.rs | 1 +
 1 file changed, 1 insertion(+)

diff --git a/crates/core_arch/src/acle/simd32.rs b/crates/core_arch/src/acle/simd32.rs
index e5d00e1451..1c64edee92 100644
--- a/crates/core_arch/src/acle/simd32.rs
+++ b/crates/core_arch/src/acle/simd32.rs
@@ -353,6 +353,7 @@ pub unsafe fn __shsub8(a: int8x4_t, b: int8x4_t) -> int8x4_t {
 /// res\[3\] = a\[3\] - a\[3\]
 ///
 /// where [0] is the lower 8 bits and [3] is the upper 8 bits.
+/// The GE bits of the APSR are set.
 #[inline]
 #[cfg_attr(test, assert_instr(usub8))]
 pub unsafe fn __usub8(a: uint8x4_t, b: uint8x4_t) -> uint8x4_t {