Skip to content

s390x: use llvm intrinsics instead of simd_fmin/fmax #2058

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
folkertdev merged 3 commits into from
Mar 11, 2026
+31 −4
Merged

s390x: use llvm intrinsics instead of simd_fmin/fmax #2058

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6efd06e
s390x: use llvm.s390 intrinsics instead of simd_fmin/fmax
RalfJung Mar 10, 2026
bfc3662
add f32 min/max tests
RalfJung Mar 10, 2026
454668b
go back to portable LLVM intrinsic to avoid fallback trouble
RalfJung Mar 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 31 additions & 4 deletions crates/core_arch/src/s390x/vector.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -335,6 +335,20 @@ unsafe extern "unadjusted" {
#[link_name = "llvm.s390.vcfn"] fn vcfn(a: vector_signed_short, immarg: i32) -> vector_signed_short;
#[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;

// 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: vector_float, b: vector_float) -> vector_float;
#[link_name = "llvm.minnum.v2f64"] fn minnum_v2f64(a: vector_double, b: vector_double) -> vector_double;
#[link_name = "llvm.maxnum.v4f32"] fn maxnum_v4f32(a: vector_float, b: vector_float) -> vector_float;
#[link_name = "llvm.maxnum.v2f64"] fn maxnum_v2f64(a: vector_double, b: vector_double) -> vector_double;
}

#[repr(simd)]
Expand Down Expand Up @@ -780,8 +794,8 @@ mod sealed {
impl_max!(vec_vmxslg, vector_unsigned_long_long, vmxlg);
}

test_impl! { vec_vfmaxsb (a: vector_float, b: vector_float) -> vector_float [simd_fmax, "vector-enhancements-1" vfmaxsb ] }
test_impl! { vec_vfmaxdb (a: vector_double, b: vector_double) -> vector_double [simd_fmax, "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);
Expand Down Expand Up @@ -827,8 +841,8 @@ mod sealed {
impl_min!(vec_vmnslg, vector_unsigned_long_long, vmnlg);
}

test_impl! { vec_vfminsb (a: vector_float, b: vector_float) -> vector_float [simd_fmin, "vector-enhancements-1" vfminsb] }
test_impl! { vec_vfmindb (a: vector_double, b: vector_double) -> vector_double [simd_fmin, "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);
Expand Down Expand Up @@ -7477,6 +7491,19 @@ mod tests {
[0, !0, !0, !0]
}

// f32 is the tricky case for max/min as that needs a fallback on z13
Copy link
Member Author

@RalfJung RalfJung Mar 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@uweigand so... this comment is wrong then?

Copy link

@uweigand uweigand Mar 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah yes. On z13 the f32 case would cause a compile-time error in the C version. The tricky case (available but emulated) is f64 on z13.

test_vec_2! { test_vec_max, vec_max, f32x4, f32x4 -> f32x4,
[1.0, f32::NAN, f32::INFINITY, 2.0],
[-10.0, -10.0, 5.0, f32::NAN],
[1.0, -10.0, f32::INFINITY, 2.0]
}

test_vec_2! { test_vec_min, vec_min, f32x4, f32x4 -> f32x4,
[1.0, f32::NAN, f32::INFINITY, 2.0],
[-10.0, -10.0, 5.0, f32::NAN],
[-10.0, -10.0, 5.0, 2.0]
}

#[simd_test(enable = "vector")]
fn test_vec_meadd() {
let a = vector_unsigned_short([1, 0, 2, 0, 3, 0, 4, 0]);
Expand Down
Loading