From e67120fc2f2cb9a355c7b717508134699aa016f1 Mon Sep 17 00:00:00 2001
From: Tobias Decking <Tobias.Decking@gmail.com>
Date: Mon, 1 Jul 2024 14:24:03 +0200
Subject: [PATCH] Fix incorrect reduction operations in avx512f

---
 crates/core_arch/src/x86/avx512f.rs | 33 +++++++++--------------------
 1 file changed, 10 insertions(+), 23 deletions(-)

diff --git a/crates/core_arch/src/x86/avx512f.rs b/crates/core_arch/src/x86/avx512f.rs
index 8c88d3aa2f..3d16a409a6 100644
--- a/crates/core_arch/src/x86/avx512f.rs
+++ b/crates/core_arch/src/x86/avx512f.rs
@@ -31571,7 +31571,7 @@ pub unsafe fn _mm512_mask_reduce_max_epi32(k: __mmask16, a: __m512i) -> i32 {
     simd_reduce_max(simd_select_bitmask(
         k,
         a.as_i32x16(),
-        _mm512_undefined_epi32().as_i32x16(),
+        i32x16::splat(i32::MIN),
     ))
 }
 
@@ -31592,11 +31592,7 @@ pub unsafe fn _mm512_reduce_max_epi64(a: __m512i) -> i64 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_max_epi64(k: __mmask8, a: __m512i) -> i64 {
-    simd_reduce_max(simd_select_bitmask(
-        k,
-        a.as_i64x8(),
-        _mm512_set1_epi64(0).as_i64x8(),
-    ))
+    simd_reduce_max(simd_select_bitmask(k, a.as_i64x8(), i64x8::splat(i64::MIN)))
 }
 
 /// Reduce the packed unsigned 32-bit integers in a by maximum. Returns the maximum of all elements in a.
@@ -31619,7 +31615,7 @@ pub unsafe fn _mm512_mask_reduce_max_epu32(k: __mmask16, a: __m512i) -> u32 {
     simd_reduce_max(simd_select_bitmask(
         k,
         a.as_u32x16(),
-        _mm512_undefined_epi32().as_u32x16(),
+        _mm512_setzero_si512().as_u32x16(),
     ))
 }
 
@@ -31643,7 +31639,7 @@ pub unsafe fn _mm512_mask_reduce_max_epu64(k: __mmask8, a: __m512i) -> u64 {
     simd_reduce_max(simd_select_bitmask(
         k,
         a.as_u64x8(),
-        _mm512_set1_epi64(0).as_u64x8(),
+        _mm512_setzero_si512().as_u64x8(),
     ))
 }
 
@@ -31718,7 +31714,7 @@ pub unsafe fn _mm512_mask_reduce_min_epi32(k: __mmask16, a: __m512i) -> i32 {
     simd_reduce_min(simd_select_bitmask(
         k,
         a.as_i32x16(),
-        _mm512_undefined_epi32().as_i32x16(),
+        i32x16::splat(0x7fffffff),
     ))
 }
 
@@ -31742,7 +31738,7 @@ pub unsafe fn _mm512_mask_reduce_min_epi64(k: __mmask8, a: __m512i) -> i64 {
     simd_reduce_min(simd_select_bitmask(
         k,
         a.as_i64x8(),
-        _mm512_set1_epi64(0).as_i64x8(),
+        i64x8::splat(0x7fffffff_ffffffff),
     ))
 }
 
@@ -31766,7 +31762,7 @@ pub unsafe fn _mm512_mask_reduce_min_epu32(k: __mmask16, a: __m512i) -> u32 {
     simd_reduce_min(simd_select_bitmask(
         k,
         a.as_u32x16(),
-        _mm512_undefined_epi32().as_u32x16(),
+        u32x16::splat(0xffffffff),
     ))
 }
 
@@ -31790,7 +31786,7 @@ pub unsafe fn _mm512_mask_reduce_min_epu64(k: __mmask8, a: __m512i) -> u64 {
     simd_reduce_min(simd_select_bitmask(
         k,
         a.as_u64x8(),
-        _mm512_set1_epi64(0).as_u64x8(),
+        u64x8::splat(0xffffffff_ffffffff),
     ))
 }
 
@@ -31862,11 +31858,7 @@ pub unsafe fn _mm512_reduce_and_epi32(a: __m512i) -> i32 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_and_epi32(k: __mmask16, a: __m512i) -> i32 {
-    simd_reduce_and(simd_select_bitmask(
-        k,
-        a.as_i32x16(),
-        _mm512_set1_epi32(0xFF).as_i32x16(),
-    ))
+    simd_reduce_and(simd_select_bitmask(k, a.as_i32x16(), i32x16::splat(-1)))
 }
 
 /// Reduce the packed 64-bit integers in a by bitwise AND. Returns the bitwise AND of all elements in a.
@@ -31886,12 +31878,7 @@ pub unsafe fn _mm512_reduce_and_epi64(a: __m512i) -> i64 {
 #[target_feature(enable = "avx512f")]
 #[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
 pub unsafe fn _mm512_mask_reduce_and_epi64(k: __mmask8, a: __m512i) -> i64 {
-    simd_reduce_and(simd_select_bitmask(
-        k,
-        a.as_i64x8(),
-        _mm512_set1_epi64(1 << 0 | 1 << 1 | 1 << 2 | 1 << 3 | 1 << 4 | 1 << 5 | 1 << 6 | 1 << 7)
-            .as_i64x8(),
-    ))
+    simd_reduce_and(simd_select_bitmask(k, a.as_i64x8(), i64x8::splat(-1)))
 }
 
 /// Reduce the packed 32-bit integers in a by bitwise OR. Returns the bitwise OR of all elements in a.