Rollup merge of #135785 - folkertdev:s390x-vector-passmode-direct, r=bjorn3

use `PassMode::Direct` for vector types on `s390x`

closes #135744
tracking issue: #130869

Previously, all vector types were type erased to `Ni8`, now we pass non-wrapped vector types directly. That skips emitting a bunch of casting logic in rustc, that LLVM then has to clean up. The initial LLVM IR is also a bit more readable.

This calling convention is tested extensively in `tests/assembly/s390x-vector-abi.rs`, showing that this change has no impact on the ABI in practice.

r? ````@taiki-e````

Author: jhpratt

compiler/rustc_target/src/callconv/s390x.rs

@@ -38,9 +38,17 @@ where
     }
 
     let size = arg.layout.size;
-    if size.bits() <= 128 && arg.layout.is_single_vector_element(cx, size) {
-        arg.cast_to(Reg { kind: RegKind::Vector, size });
-        return;
+    if size.bits() <= 128 {
+        if let BackendRepr::Vector { .. } = arg.layout.backend_repr {
+            // pass non-wrapped vector types using `PassMode::Direct`
+            return;
+        }
+
+        if arg.layout.is_single_vector_element(cx, size) {
+            // pass non-transparant wrappers around a vector as `PassMode::Cast`
+            arg.cast_to(Reg { kind: RegKind::Vector, size });
+            return;
+        }
     }
     if !arg.layout.is_aggregate() && size.bits() <= 64 {
         arg.extend_integer_width_to(64);

tests/codegen/s390x-simd.rs

@@ -0,0 +1,143 @@
+//! test that s390x vector types are passed using `PassMode::Direct`
+//! see also https://github.com/rust-lang/rust/issues/135744
+//@ add-core-stubs
+//@ compile-flags: --target s390x-unknown-linux-gnu -O
+//@ needs-llvm-components: systemz
+
+#![crate_type = "rlib"]
+#![feature(no_core, asm_experimental_arch)]
+#![feature(s390x_target_feature, simd_ffi, link_llvm_intrinsics, repr_simd)]
+#![no_core]
+
+extern crate minicore;
+use minicore::*;
+
+#[repr(simd)]
+struct i8x16([i8; 16]);
+
+#[repr(simd)]
+struct i16x8([i16; 8]);
+
+#[repr(simd)]
+struct i32x4([i32; 4]);
+
+#[repr(simd)]
+struct i64x2([i64; 2]);
+
+#[repr(simd)]
+struct f32x4([f32; 4]);
+
+#[repr(simd)]
+struct f64x2([f64; 2]);
+
+#[allow(improper_ctypes)]
+extern "C" {
+    #[link_name = "llvm.smax.v16i8"]
+    fn vmxb(a: i8x16, b: i8x16) -> i8x16;
+    #[link_name = "llvm.smax.v8i16"]
+    fn vmxh(a: i16x8, b: i16x8) -> i16x8;
+    #[link_name = "llvm.smax.v4i32"]
+    fn vmxf(a: i32x4, b: i32x4) -> i32x4;
+    #[link_name = "llvm.smax.v2i64"]
+    fn vmxg(a: i64x2, b: i64x2) -> i64x2;
+}
+
+// CHECK-LABEL: define <16 x i8> @max_i8x16
+// CHECK-SAME: <16 x i8> %a, <16 x i8> %b
+// CHECK: call <16 x i8> @llvm.smax.v16i8(<16 x i8> %a, <16 x i8> %b)
+#[no_mangle]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn max_i8x16(a: i8x16, b: i8x16) -> i8x16 {
+    vmxb(a, b)
+}
+
+// CHECK-LABEL: define <8 x i16> @max_i16x8
+// CHECK-SAME: <8 x i16> %a, <8 x i16> %b
+// CHECK: call <8 x i16> @llvm.smax.v8i16(<8 x i16> %a, <8 x i16> %b)
+#[no_mangle]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn max_i16x8(a: i16x8, b: i16x8) -> i16x8 {
+    vmxh(a, b)
+}
+
+// CHECK-LABEL: define <4 x i32> @max_i32x4
+// CHECK-SAME: <4 x i32> %a, <4 x i32> %b
+// CHECK: call <4 x i32> @llvm.smax.v4i32(<4 x i32> %a, <4 x i32> %b)
+#[no_mangle]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn max_i32x4(a: i32x4, b: i32x4) -> i32x4 {
+    vmxf(a, b)
+}
+
+// CHECK-LABEL: define <2 x i64> @max_i64x2
+// CHECK-SAME: <2 x i64> %a, <2 x i64> %b
+// CHECK: call <2 x i64> @llvm.smax.v2i64(<2 x i64> %a, <2 x i64> %b)
+#[no_mangle]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn max_i64x2(a: i64x2, b: i64x2) -> i64x2 {
+    vmxg(a, b)
+}
+
+// CHECK-LABEL: define <4 x float> @choose_f32x4
+// CHECK-SAME: <4 x float> %a, <4 x float> %b
+#[no_mangle]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn choose_f32x4(a: f32x4, b: f32x4, c: bool) -> f32x4 {
+    if c { a } else { b }
+}
+
+// CHECK-LABEL: define <2 x double> @choose_f64x2
+// CHECK-SAME: <2 x double> %a, <2 x double> %b
+#[no_mangle]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn choose_f64x2(a: f64x2, b: f64x2, c: bool) -> f64x2 {
+    if c { a } else { b }
+}
+
+#[repr(C)]
+struct Wrapper<T>(T);
+
+#[no_mangle]
+#[inline(never)]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn max_wrapper_i8x16(a: Wrapper<i8x16>, b: Wrapper<i8x16>) -> Wrapper<i8x16> {
+    // CHECK-LABEL: max_wrapper_i8x16
+    // CHECK-SAME: sret([16 x i8])
+    // CHECK-SAME: <16 x i8>
+    // CHECK-SAME: <16 x i8>
+    // CHECK: call <16 x i8> @llvm.smax.v16i8
+    // CHECK-SAME: <16 x i8>
+    // CHECK-SAME: <16 x i8>
+    Wrapper(vmxb(a.0, b.0))
+}
+
+#[no_mangle]
+#[inline(never)]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn max_wrapper_i64x2(a: Wrapper<i64x2>, b: Wrapper<i64x2>) -> Wrapper<i64x2> {
+    // CHECK-LABEL: max_wrapper_i64x2
+    // CHECK-SAME: sret([16 x i8])
+    // CHECK-SAME: <16 x i8>
+    // CHECK-SAME: <16 x i8>
+    // CHECK: call <2 x i64> @llvm.smax.v2i64
+    // CHECK-SAME: <2 x i64>
+    // CHECK-SAME: <2 x i64>
+    Wrapper(vmxg(a.0, b.0))
+}
+
+#[no_mangle]
+#[inline(never)]
+#[target_feature(enable = "vector")]
+pub unsafe extern "C" fn choose_wrapper_f64x2(
+    a: Wrapper<f64x2>,
+    b: Wrapper<f64x2>,
+    c: bool,
+) -> Wrapper<f64x2> {
+    // CHECK-LABEL: choose_wrapper_f64x2
+    // CHECK-SAME: sret([16 x i8])
+    // CHECK-SAME: <16 x i8>
+    // CHECK-SAME: <16 x i8>
+    Wrapper(choose_f64x2(a.0, b.0, c))
+}
+
+// CHECK: declare <2 x i64> @llvm.smax.v2i64(<2 x i64>, <2 x i64>)