From 387c45ae167dbb4d4cfcf7590bf98e060400e775 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Eduardo=20S=C3=A1nchez=20Mu=C3=B1oz?= <eduardosm-dev@e64.io>
Date: Tue, 3 Oct 2023 23:14:09 +0200
Subject: [PATCH] Reimplement _mm_avg_epu8, _mm_avg_epu16, _mm256_avg_epu8 and
 _mm256_avg_epu16 without LLVM intrinsics

---
 library/stdarch/crates/core_arch/src/x86/avx2.rs | 14 ++++++++------
 library/stdarch/crates/core_arch/src/x86/sse2.rs | 14 ++++++++------
 2 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/library/stdarch/crates/core_arch/src/x86/avx2.rs b/library/stdarch/crates/core_arch/src/x86/avx2.rs
index 93cea66aa4e7..80c59fe746b1 100644
--- a/library/stdarch/crates/core_arch/src/x86/avx2.rs
+++ b/library/stdarch/crates/core_arch/src/x86/avx2.rs
@@ -344,7 +344,10 @@ pub unsafe fn _mm256_andnot_si256(a: __m256i, b: __m256i) -> __m256i {
 #[cfg_attr(test, assert_instr(vpavgw))]
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm256_avg_epu16(a: __m256i, b: __m256i) -> __m256i {
-    transmute(pavgw(a.as_u16x16(), b.as_u16x16()))
+    let a = simd_cast::<_, u32x16>(a.as_u16x16());
+    let b = simd_cast::<_, u32x16>(b.as_u16x16());
+    let r = simd_shr(simd_add(simd_add(a, b), u32x16::splat(1)), u32x16::splat(1));
+    transmute(simd_cast::<_, u16x16>(r))
 }
 
 /// Averages packed unsigned 8-bit integers in `a` and `b`.
@@ -355,7 +358,10 @@ pub unsafe fn _mm256_avg_epu16(a: __m256i, b: __m256i) -> __m256i {
 #[cfg_attr(test, assert_instr(vpavgb))]
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm256_avg_epu8(a: __m256i, b: __m256i) -> __m256i {
-    transmute(pavgb(a.as_u8x32(), b.as_u8x32()))
+    let a = simd_cast::<_, u16x32>(a.as_u8x32());
+    let b = simd_cast::<_, u16x32>(b.as_u8x32());
+    let r = simd_shr(simd_add(simd_add(a, b), u16x32::splat(1)), u16x32::splat(1));
+    transmute(simd_cast::<_, u8x32>(r))
 }
 
 /// Blends packed 32-bit integers from `a` and `b` using control mask `IMM4`.
@@ -3638,10 +3644,6 @@ extern "C" {
     fn pabsw(a: i16x16) -> u16x16;
     #[link_name = "llvm.x86.avx2.pabs.d"]
     fn pabsd(a: i32x8) -> u32x8;
-    #[link_name = "llvm.x86.avx2.pavg.b"]
-    fn pavgb(a: u8x32, b: u8x32) -> u8x32;
-    #[link_name = "llvm.x86.avx2.pavg.w"]
-    fn pavgw(a: u16x16, b: u16x16) -> u16x16;
     #[link_name = "llvm.x86.avx2.pblendvb"]
     fn pblendvb(a: i8x32, b: i8x32, mask: i8x32) -> i8x32;
     #[link_name = "llvm.x86.avx2.phadd.w"]
diff --git a/library/stdarch/crates/core_arch/src/x86/sse2.rs b/library/stdarch/crates/core_arch/src/x86/sse2.rs
index 107f0b0cf206..22fa62ed7d3c 100644
--- a/library/stdarch/crates/core_arch/src/x86/sse2.rs
+++ b/library/stdarch/crates/core_arch/src/x86/sse2.rs
@@ -165,7 +165,10 @@ pub unsafe fn _mm_adds_epu16(a: __m128i, b: __m128i) -> __m128i {
 #[cfg_attr(test, assert_instr(pavgb))]
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_avg_epu8(a: __m128i, b: __m128i) -> __m128i {
-    transmute(pavgb(a.as_u8x16(), b.as_u8x16()))
+    let a = simd_cast::<_, u16x16>(a.as_u8x16());
+    let b = simd_cast::<_, u16x16>(b.as_u8x16());
+    let r = simd_shr(simd_add(simd_add(a, b), u16x16::splat(1)), u16x16::splat(1));
+    transmute(simd_cast::<_, u8x16>(r))
 }
 
 /// Averages packed unsigned 16-bit integers in `a` and `b`.
@@ -176,7 +179,10 @@ pub unsafe fn _mm_avg_epu8(a: __m128i, b: __m128i) -> __m128i {
 #[cfg_attr(test, assert_instr(pavgw))]
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_avg_epu16(a: __m128i, b: __m128i) -> __m128i {
-    transmute(pavgw(a.as_u16x8(), b.as_u16x8()))
+    let a = simd_cast::<_, u32x8>(a.as_u16x8());
+    let b = simd_cast::<_, u32x8>(b.as_u16x8());
+    let r = simd_shr(simd_add(simd_add(a, b), u32x8::splat(1)), u32x8::splat(1));
+    transmute(simd_cast::<_, u16x8>(r))
 }
 
 /// Multiplies and then horizontally add signed 16 bit integers in `a` and `b`.
@@ -2838,10 +2844,6 @@ extern "C" {
     fn lfence();
     #[link_name = "llvm.x86.sse2.mfence"]
     fn mfence();
-    #[link_name = "llvm.x86.sse2.pavg.b"]
-    fn pavgb(a: u8x16, b: u8x16) -> u8x16;
-    #[link_name = "llvm.x86.sse2.pavg.w"]
-    fn pavgw(a: u16x8, b: u16x8) -> u16x8;
     #[link_name = "llvm.x86.sse2.pmadd.wd"]
     fn pmaddwd(a: i16x8, b: i16x8) -> i32x4;
     #[link_name = "llvm.x86.sse2.psad.bw"]