go back to portable LLVM intrinsic to avoid fallback trouble

Author: RalfJung

crates/core_arch/src/s390x/vector.rs

@@ -336,10 +336,19 @@ unsafe extern "unadjusted" {
     #[link_name = "llvm.s390.vcnf"] fn vcnf(a: vector_signed_short, immarg: i32) -> vector_signed_short;
     #[link_name = "llvm.s390.vcrnfs"] fn vcrnfs(a: vector_float, b: vector_float, immarg: i32) -> vector_signed_short;
 
-    #[link_name = "llvm.s390.vfmaxsb"] fn vfmaxsb(a: vector_float, b: vector_float, mode: i32) -> vector_float;
-    #[link_name = "llvm.s390.vfmaxdb"] fn vfmaxdb(a: vector_double, b: vector_double, mode: i32) -> vector_double;
-    #[link_name = "llvm.s390.vfminsb"] fn vfminsb(a: vector_float, b: vector_float, mode: i32) -> vector_float;
-    #[link_name = "llvm.s390.vfmindb"] fn vfmindb(a: vector_double, b: vector_double, mode: i32) -> vector_double;
+    // These are the intrinsics we'd like to use (with mode 0). However, they require
+    // "vector-enhancements-1" and don't have a fallback, whereas `vec_min`/`vec_max` should be
+    // available with just "vector". Therefore, we cannot use them.
+    // #[link_name = "llvm.s390.vfmaxsb"] fn vfmaxsb(a: vector_float, b: vector_float, mode: i32) -> vector_float;
+    // #[link_name = "llvm.s390.vfmaxdb"] fn vfmaxdb(a: vector_double, b: vector_double, mode: i32) -> vector_double;
+    // #[link_name = "llvm.s390.vfminsb"] fn vfminsb(a: vector_float, b: vector_float, mode: i32) -> vector_float;
+    // #[link_name = "llvm.s390.vfmindb"] fn vfmindb(a: vector_double, b: vector_double, mode: i32) -> vector_double;
+    // Instead, we use "portable" LLVM intrinsics -- even though those have the wrong semantics
+    // (https://github.com/rust-lang/stdarch/issues/2060), they usually do the right thing.
+    #[link_name = "llvm.minnum.v4f32"] fn minnum_v4f32(a: f32x4, b: f32x4) -> f32x4;
+    #[link_name = "llvm.minnum.v2f64"] fn minnum_v2f64(a: f64x2, b: f64x2) -> f64x2;
+    #[link_name = "llvm.maxnum.v4f32"] fn maxnum_v4f32(a: f32x4, b: f32x4) -> f32x4;
+    #[link_name = "llvm.maxnum.v2f64"] fn maxnum_v2f64(a: f64x2, b: f64x2) -> f64x2;
 }
 
 #[repr(simd)]
@@ -797,8 +806,8 @@ mod sealed {
         vfmaxdb(a, b, const { 0 })
     }
 
-    test_impl! { vec_vfmaxsb (a: vector_float, b: vector_float) -> vector_float [vfmaxsb_m0, "vector-enhancements-1" vfmaxsb ] }
-    test_impl! { vec_vfmaxdb (a: vector_double, b: vector_double) -> vector_double [vfmaxdb_m0, "vector-enhancements-1" vfmaxdb] }
+    test_impl! { vec_vfmaxsb (a: vector_float, b: vector_float) -> vector_float [maxnum_v4f32, "vector-enhancements-1" vfmaxsb]  }
+    test_impl! { vec_vfmaxdb (a: vector_double, b: vector_double) -> vector_double [maxnum_v2f64, "vector-enhancements-1" vfmaxdb]  }
 
     impl_vec_trait!([VectorMax vec_max] vec_vfmaxsb (vector_float, vector_float) -> vector_float);
     impl_vec_trait!([VectorMax vec_max] vec_vfmaxdb (vector_double, vector_double) -> vector_double);
@@ -844,20 +853,8 @@ mod sealed {
         impl_min!(vec_vmnslg, vector_unsigned_long_long, vmnlg);
     }
 
-    #[inline]
-    #[target_feature(enable = "vector")]
-    unsafe fn vfminsb_m0(a: vector_float, b: vector_float) -> vector_float {
-        // clang uses mode 0 for `vec_min`, so we do the same.
-        vfminsb(a, b, const { 0 })
-    }
-    #[inline]
-    #[target_feature(enable = "vector")]
-    unsafe fn vfmindb_m0(a: vector_double, b: vector_double) -> vector_double {
-        vfmindb(a, b, const { 0 })
-    }
-
-    test_impl! { vec_vfminsb (a: vector_float, b: vector_float) -> vector_float [vfminsb_m0, "vector-enhancements-1" vfminsb]  }
-    test_impl! { vec_vfmindb (a: vector_double, b: vector_double) -> vector_double [vfmindb_m0, "vector-enhancements-1" vfmindb]  }
+    test_impl! { vec_vfminsb (a: vector_float, b: vector_float) -> vector_float [minnum_v4f32, "vector-enhancements-1" vfminsb]  }
+    test_impl! { vec_vfmindb (a: vector_double, b: vector_double) -> vector_double [minnum_v2f64, "vector-enhancements-1" vfmindb]  }
 
     impl_vec_trait!([VectorMin vec_min] vec_vfminsb (vector_float, vector_float) -> vector_float);
     impl_vec_trait!([VectorMin vec_min] vec_vfmindb (vector_double, vector_double) -> vector_double);