Rollup merge of #133102 - RalfJung:aarch64-softfloat, r=davidtwco,wesleywiser

aarch64 softfloat target: always pass floats in int registers

This is a part of #131058: on softfloat aarch64 targets, the float registers may be unavailable. And yet, LLVM will happily use them to pass float types if the corresponding target features are enabled. That's a problem as it means enabling/disabling `neon` instructions can change the ABI.

Other targets have a `soft-float` target feature that forces the use of the soft-float ABI no matter whether float registers are enabled or not; aarch64 has nothing like that.

So we follow the aarch64 [softfloat ABI](#131058 (comment)) and treat floats like integers for `extern "C"` functions. For the "Rust" ABI, we do the same for scalars, and then just do something reasonable for ScalarPair that avoids the pointer indirection.

Cc ```@workingjubilee```

Author: compiler-errors

compiler/rustc_target/src/callconv/aarch64.rs

@@ -1,5 +1,10 @@
+use std::iter;
+
+use rustc_abi::{BackendRepr, Primitive};
+
 use crate::abi::call::{ArgAbi, FnAbi, Reg, RegKind, Uniform};
 use crate::abi::{HasDataLayout, TyAbiInterface};
+use crate::spec::{HasTargetSpec, Target};
 
 /// Indicates the variant of the AArch64 ABI we are compiling for.
 /// Used to accommodate Apple and Microsoft's deviations from the usual AAPCS ABI.
@@ -15,7 +20,7 @@ pub(crate) enum AbiKind {
 fn is_homogeneous_aggregate<'a, Ty, C>(cx: &C, arg: &mut ArgAbi<'a, Ty>) -> Option<Uniform>
 where
     Ty: TyAbiInterface<'a, C> + Copy,
-    C: HasDataLayout,
+    C: HasDataLayout + HasTargetSpec,
 {
     arg.layout.homogeneous_aggregate(cx).ok().and_then(|ha| ha.unit()).and_then(|unit| {
         let size = arg.layout.size;
@@ -27,18 +32,52 @@ where
 
         let valid_unit = match unit.kind {
             RegKind::Integer => false,
-            RegKind::Float => true,
+            // The softfloat ABI treats floats like integers, so they
+            // do not get homogeneous aggregate treatment.
+            RegKind::Float => cx.target_spec().abi != "softfloat",
             RegKind::Vector => size.bits() == 64 || size.bits() == 128,
         };
 
         valid_unit.then_some(Uniform::consecutive(unit, size))
     })
 }
 
+fn softfloat_float_abi<Ty>(target: &Target, arg: &mut ArgAbi<'_, Ty>) {
+    if target.abi != "softfloat" {
+        return;
+    }
+    // Do *not* use the float registers for passing arguments, as that would make LLVM pick the ABI
+    // and its choice depends on whether `neon` instructions are enabled. Instead, we follow the
+    // AAPCS "softfloat" ABI, which specifies that floats should be passed as equivalently-sized
+    // integers. Nominally this only exists for "R" profile chips, but sometimes people don't want
+    // to use hardfloats even if the hardware supports them, so we do this for all softfloat
+    // targets.
+    if let BackendRepr::Scalar(s) = arg.layout.backend_repr
+        && let Primitive::Float(f) = s.primitive()
+    {
+        arg.cast_to(Reg { kind: RegKind::Integer, size: f.size() });
+    } else if let BackendRepr::ScalarPair(s1, s2) = arg.layout.backend_repr
+        && (matches!(s1.primitive(), Primitive::Float(_))
+            || matches!(s2.primitive(), Primitive::Float(_)))
+    {
+        // This case can only be reached for the Rust ABI, so we can do whatever we want here as
+        // long as it does not depend on target features (i.e., as long as we do not use float
+        // registers). So we pass small things in integer registers and large things via pointer
+        // indirection. This means we lose the nice "pass it as two arguments" optimization, but we
+        // currently just have to way to combine a `PassMode::Cast` with that optimization (and we
+        // need a cast since we want to pass the float as an int).
+        if arg.layout.size.bits() <= target.pointer_width.into() {
+            arg.cast_to(Reg { kind: RegKind::Integer, size: arg.layout.size });
+        } else {
+            arg.make_indirect();
+        }
+    }
+}
+
 fn classify_ret<'a, Ty, C>(cx: &C, ret: &mut ArgAbi<'a, Ty>, kind: AbiKind)
 where
     Ty: TyAbiInterface<'a, C> + Copy,
-    C: HasDataLayout,
+    C: HasDataLayout + HasTargetSpec,
 {
     if !ret.layout.is_sized() {
         // Not touching this...
@@ -51,6 +90,7 @@ where
             // See also: <https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms#Pass-Arguments-to-Functions-Correctly>
             ret.extend_integer_width_to(32)
         }
+        softfloat_float_abi(cx.target_spec(), ret);
         return;
     }
     if let Some(uniform) = is_homogeneous_aggregate(cx, ret) {
@@ -69,7 +109,7 @@ where
 fn classify_arg<'a, Ty, C>(cx: &C, arg: &mut ArgAbi<'a, Ty>, kind: AbiKind)
 where
     Ty: TyAbiInterface<'a, C> + Copy,
-    C: HasDataLayout,
+    C: HasDataLayout + HasTargetSpec,
 {
     if !arg.layout.is_sized() {
         // Not touching this...
@@ -82,6 +122,8 @@ where
             // See also: <https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms#Pass-Arguments-to-Functions-Correctly>
             arg.extend_integer_width_to(32);
         }
+        softfloat_float_abi(cx.target_spec(), arg);
+
         return;
     }
     if let Some(uniform) = is_homogeneous_aggregate(cx, arg) {
@@ -112,7 +154,7 @@ where
 pub(crate) fn compute_abi_info<'a, Ty, C>(cx: &C, fn_abi: &mut FnAbi<'a, Ty>, kind: AbiKind)
 where
     Ty: TyAbiInterface<'a, C> + Copy,
-    C: HasDataLayout,
+    C: HasDataLayout + HasTargetSpec,
 {
     if !fn_abi.ret.is_ignore() {
         classify_ret(cx, &mut fn_abi.ret, kind);
@@ -125,3 +167,13 @@ where
         classify_arg(cx, arg, kind);
     }
 }
+
+pub(crate) fn compute_rust_abi_info<'a, Ty, C>(cx: &C, fn_abi: &mut FnAbi<'a, Ty>)
+where
+    Ty: TyAbiInterface<'a, C> + Copy,
+    C: HasDataLayout + HasTargetSpec,
+{
+    for arg in fn_abi.args.iter_mut().chain(iter::once(&mut fn_abi.ret)) {
+        softfloat_float_abi(cx.target_spec(), arg);
+    }
+}

compiler/rustc_target/src/callconv/mod.rs

@@ -738,6 +738,7 @@ impl<'a, Ty> FnAbi<'a, Ty> {
             "x86" => x86::compute_rust_abi_info(cx, self, abi),
             "riscv32" | "riscv64" => riscv::compute_rust_abi_info(cx, self, abi),
             "loongarch64" => loongarch::compute_rust_abi_info(cx, self, abi),
+            "aarch64" => aarch64::compute_rust_abi_info(cx, self),
             _ => {}
         };
 

tests/codegen/aarch64-softfloat.rs

@@ -0,0 +1,48 @@
+//@ compile-flags: --target aarch64-unknown-none-softfloat -Zmerge-functions=disabled
+//@ needs-llvm-components: aarch64
+#![crate_type = "lib"]
+#![feature(no_core, lang_items)]
+#![no_core]
+
+#[lang = "sized"]
+trait Sized {}
+#[lang = "copy"]
+trait Copy {}
+impl Copy for f32 {}
+impl Copy for f64 {}
+
+// CHECK: i64 @pass_f64_C(i64 {{[^,]*}})
+#[no_mangle]
+extern "C" fn pass_f64_C(x: f64) -> f64 {
+    x
+}
+
+// CHECK: i64 @pass_f32_pair_C(i64 {{[^,]*}})
+#[no_mangle]
+extern "C" fn pass_f32_pair_C(x: (f32, f32)) -> (f32, f32) {
+    x
+}
+
+// CHECK: [2 x i64] @pass_f64_pair_C([2 x i64] {{[^,]*}})
+#[no_mangle]
+extern "C" fn pass_f64_pair_C(x: (f64, f64)) -> (f64, f64) {
+    x
+}
+
+// CHECK: i64 @pass_f64_Rust(i64 {{[^,]*}})
+#[no_mangle]
+fn pass_f64_Rust(x: f64) -> f64 {
+    x
+}
+
+// CHECK: i64 @pass_f32_pair_Rust(i64 {{[^,]*}})
+#[no_mangle]
+fn pass_f32_pair_Rust(x: (f32, f32)) -> (f32, f32) {
+    x
+}
+
+// CHECK: void @pass_f64_pair_Rust(ptr {{[^,]*}}, ptr {{[^,]*}})
+#[no_mangle]
+fn pass_f64_pair_Rust(x: (f64, f64)) -> (f64, f64) {
+    x
+}