From d099f72175907fc5f8346e2e06daba88bb34cf12 Mon Sep 17 00:00:00 2001
From: Lokathor <zefria@gmail.com>
Date: Sun, 11 Aug 2024 23:04:05 -0600
Subject: [PATCH] Sse4 fix (#178)

* restore a test that wasn't marked as a test.

* remove the case that calls non-existing functions

* add a CI to prevent this from happening again.

* conditional sse4.1 testing
---
 .github/workflows/rust.yml | 12 ++++++++----
 src/u16x16_.rs             |  7 +------
 tests/all_tests/t_u32x8.rs |  3 ++-
 3 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/rust.yml b/.github/workflows/rust.yml
index 20db88f..c290b9a 100644
--- a/.github/workflows/rust.yml
+++ b/.github/workflows/rust.yml
@@ -21,10 +21,10 @@ jobs:
         - { target: i686-pc-windows-msvc, toolchain: beta, os: windows-latest }
         - { target: i686-pc-windows-msvc, toolchain: nightly, os: windows-latest }
         # x86_64
-        - { target: x86_64-pc-windows-msvc, toolchain: "1.61", os: windows-latest }
-        - { target: x86_64-pc-windows-msvc, toolchain: stable, os: windows-latest }
-        - { target: x86_64-pc-windows-msvc, toolchain: beta, os: windows-latest }
-        - { target: x86_64-pc-windows-msvc, toolchain: nightly, os: windows-latest }
+        - { target: x86_64-unknown-linux-gnu, toolchain: "1.61", os: ubuntu-latest }
+        - { target: x86_64-unknown-linux-gnu, toolchain: stable, os: ubuntu-latest }
+        - { target: x86_64-unknown-linux-gnu, toolchain: beta, os: ubuntu-latest }
+        - { target: x86_64-unknown-linux-gnu, toolchain: nightly, os: ubuntu-latest }
         # aarch64
         - { target: aarch64-apple-darwin, toolchain: "1.61", os: macos-latest }
         - { target: aarch64-apple-darwin, toolchain: stable, os: macos-latest }
@@ -59,6 +59,10 @@ jobs:
       env:
         CARGO_TARGET_WASM32_WASI_RUNNER: wasmtime run --wasm-features all --dir .
       run: cargo test --target ${{ matrix.rust.target }} --all-features
+    
+    - name: Build the crate with SSE4.1 (the "native" of CI will be above this)
+      if: matrix.rust.os == 'ubuntu-latest' && matrix.rust.target == 'x86_64-unknown-linux-gnu'
+      run: RUSTFLAGS="-Ctarget-feature=+sse4.1" cargo build --target ${{ matrix.rust.target }}
 
     - name: switch over to native cpu features
       run: mv .cargo-ci .cargo
diff --git a/src/u16x16_.rs b/src/u16x16_.rs
index 0ff6024..565dce8 100644
--- a/src/u16x16_.rs
+++ b/src/u16x16_.rs
@@ -231,18 +231,13 @@ impl Mul for u16x16 {
 }
 
 impl From<u8x16> for u16x16 {
-  /// widens and sign extends to i16x16
+  /// widens and sign extends to u16x16
   #[inline]
   #[must_use]
   fn from(v: u8x16) -> Self {
     pick! {
       if #[cfg(target_feature="avx2")] {
         u16x16 { avx2:convert_to_i16_m256i_from_u8_m128i(v.sse) }
-      } else if #[cfg(target_feature="sse4.1")] {
-        u16x16 {
-          a: u16x8 { sse: convert_to_i16_m128i_from_lower8_u8_m128i(v.sse) },
-          b: u16x8 { sse: convert_to_i16_m128i_from_lower8_u8_m128i(unpack_high_i64_m128i(v.sse, v.sse)) }
-        }
       } else if #[cfg(target_feature="sse2")] {
         u16x16 {
           a: u16x8 { sse: shr_imm_u16_m128i::<8>( unpack_low_i8_m128i(v.sse, v.sse)) },
diff --git a/tests/all_tests/t_u32x8.rs b/tests/all_tests/t_u32x8.rs
index d4038d4..da6599a 100644
--- a/tests/all_tests/t_u32x8.rs
+++ b/tests/all_tests/t_u32x8.rs
@@ -268,7 +268,8 @@ fn impl_u32x8_from_u16x8() {
     |a, _b| a as u32,
   );
 }
-  
+
+#[test]
 fn test_u32x8_any() {
   let a = u32x8::from([0, 0, 0, u32::MAX, 0, 0, 0, 0]);
   assert!(a.any());