Support s390x z13 vector ABI

Author: taiki-e

compiler/rustc_target/src/callconv/s390x.rs

@@ -1,16 +1,39 @@
-// FIXME: The assumes we're using the non-vector ABI, i.e., compiling
-// for a pre-z13 machine or using -mno-vx.
+// Reference: ELF Application Binary Interface s390x Supplement
+// https://github.com/IBM/s390x-abi
 
-use crate::abi::call::{ArgAbi, FnAbi, Reg};
-use crate::abi::{HasDataLayout, TyAbiInterface};
+use rustc_abi::{BackendRepr, HasDataLayout, Size, TyAbiInterface, TyAndLayout};
+
+use crate::abi::call::{ArgAbi, FnAbi, Reg, RegKind};
 use crate::spec::HasTargetSpec;
 
+fn contains_vector<'a, Ty, C>(cx: &C, layout: TyAndLayout<'a, Ty>, expected_size: Size) -> bool
+where
+    Ty: TyAbiInterface<'a, C> + Copy,
+{
+    match layout.backend_repr {
+        BackendRepr::Uninhabited | BackendRepr::Scalar(_) | BackendRepr::ScalarPair(..) => false,
+        BackendRepr::Vector { .. } => layout.size == expected_size,
+        BackendRepr::Memory { .. } => {
+            if layout.fields.count() != 1 {
+                // Zero-sized members in C are not ignored.
+                // See https://github.com/rust-lang/rust/pull/131586#discussion_r1836718798.
+                return false;
+            }
+            contains_vector(cx, layout.field(cx, 0), expected_size)
+        }
+    }
+}
+
 fn classify_ret<Ty>(ret: &mut ArgAbi<'_, Ty>) {
-    if !ret.layout.is_aggregate() && ret.layout.size.bits() <= 64 {
+    let size = ret.layout.size;
+    if size.bits() <= 128 && matches!(ret.layout.backend_repr, BackendRepr::Vector { .. }) {
+        return;
+    }
+    if !ret.layout.is_aggregate() && size.bits() <= 64 {
         ret.extend_integer_width_to(64);
-    } else {
-        ret.make_indirect();
+        return;
     }
+    ret.make_indirect();
 }
 
 fn classify_arg<'a, Ty, C>(cx: &C, arg: &mut ArgAbi<'a, Ty>)
@@ -32,19 +55,25 @@ where
         }
         return;
     }
-    if !arg.layout.is_aggregate() && arg.layout.size.bits() <= 64 {
+
+    let size = arg.layout.size;
+    if size.bits() <= 128 && contains_vector(cx, arg.layout, size) {
+        arg.cast_to(Reg { kind: RegKind::Vector, size });
+        return;
+    }
+    if !arg.layout.is_aggregate() && size.bits() <= 64 {
         arg.extend_integer_width_to(64);
         return;
     }
 
     if arg.layout.is_single_fp_element(cx) {
-        match arg.layout.size.bytes() {
+        match size.bytes() {
             4 => arg.cast_to(Reg::f32()),
             8 => arg.cast_to(Reg::f64()),
             _ => arg.make_indirect(),
         }
     } else {
-        match arg.layout.size.bytes() {
+        match size.bytes() {
             1 => arg.cast_to(Reg::i8()),
             2 => arg.cast_to(Reg::i16()),
             4 => arg.cast_to(Reg::i32()),

compiler/rustc_target/src/spec/targets/s390x_unknown_linux_gnu.rs

@@ -6,9 +6,6 @@ pub(crate) fn target() -> Target {
     base.endian = Endian::Big;
     // z10 is the oldest CPU supported by LLVM
     base.cpu = "z10".into();
-    // FIXME: The ABI implementation in abi/call/s390x.rs is for now hard-coded to assume the no-vector
-    // ABI. Pass the -vector feature string to LLVM to respect this assumption.
-    base.features = "-vector".into();
     base.max_atomic_width = Some(128);
     base.min_global_align = Some(16);
     base.stack_probes = StackProbeType::Inline;

compiler/rustc_target/src/spec/targets/s390x_unknown_linux_musl.rs

@@ -6,9 +6,6 @@ pub(crate) fn target() -> Target {
     base.endian = Endian::Big;
     // z10 is the oldest CPU supported by LLVM
     base.cpu = "z10".into();
-    // FIXME: The ABI implementation in abi/call/s390x.rs is for now hard-coded to assume the no-vector
-    // ABI. Pass the -vector feature string to LLVM to respect this assumption.
-    base.features = "-vector".into();
     base.max_atomic_width = Some(128);
     base.min_global_align = Some(16);
     base.static_position_independent_executables = true;

compiler/rustc_target/src/target_features.rs

@@ -582,6 +582,7 @@ pub fn all_rust_features() -> impl Iterator<Item = (&'static str, Stability)> {
 const X86_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] =
     &[(128, "sse"), (256, "avx"), (512, "avx512f")];
 const AARCH64_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(128, "neon")];
+const S390X_FEATURES_FOR_CORRECT_VECTOR_ABI: &'static [(u64, &'static str)] = &[(128, "vector")];
 
 impl super::spec::Target {
     pub fn rust_target_features(&self) -> &'static [(&'static str, Stability, ImpliedFeatures)] {
@@ -608,6 +609,7 @@ impl super::spec::Target {
         match &*self.arch {
             "x86" | "x86_64" => Some(X86_FEATURES_FOR_CORRECT_VECTOR_ABI),
             "aarch64" => Some(AARCH64_FEATURES_FOR_CORRECT_VECTOR_ABI),
+            "s390x" => Some(S390X_FEATURES_FOR_CORRECT_VECTOR_ABI),
             // FIXME: add support for non-tier1 architectures
             _ => None,
         }