From: WANG Rui <wangrui@loongson.cn>
Date: Thu, 21 Aug 2025 14:15:08 +0800
Subject: Fix LoongArch C function ABI when passing/returning structs
 containing floats

(cherry picked from commit b65a177b63237398dc3f49cb5598b16bb9318136)

Forwarded: not-needed
---
 compiler/rustc_target/src/callconv/loongarch.rs    | 137 ++++++++++++++++-----
 tests/assembly/loongarch-float-struct-abi.rs       | 134 ++++++++++++++++++++
 tests/codegen/cast-target-abi.rs                   |  52 ++++++--
 .../loongarch-abi/cast-local-large-enough.rs       |  44 +++++++
 4 files changed, 327 insertions(+), 40 deletions(-)
 create mode 100644 tests/assembly/loongarch-float-struct-abi.rs
 create mode 100644 tests/codegen/loongarch-abi/cast-local-large-enough.rs

diff --git a/compiler/rustc_target/src/callconv/loongarch.rs b/compiler/rustc_target/src/callconv/loongarch.rs
index c779720..76f939f 100644
--- a/compiler/rustc_target/src/callconv/loongarch.rs
+++ b/compiler/rustc_target/src/callconv/loongarch.rs
@@ -8,16 +8,16 @@ use crate::spec::HasTargetSpec;
 
 #[derive(Copy, Clone)]
 enum RegPassKind {
-    Float(Reg),
-    Integer(Reg),
+    Float { offset_from_start: Size, ty: Reg },
+    Integer { offset_from_start: Size, ty: Reg },
     Unknown,
 }
 
 #[derive(Copy, Clone)]
 enum FloatConv {
-    FloatPair(Reg, Reg),
+    FloatPair { first_ty: Reg, second_ty_offset_from_start: Size, second_ty: Reg },
     Float(Reg),
-    MixedPair(Reg, Reg),
+    MixedPair { first_ty: Reg, second_ty_offset_from_start: Size, second_ty: Reg },
 }
 
 #[derive(Copy, Clone)]
@@ -37,6 +37,7 @@ fn should_use_fp_conv_helper<'a, Ty, C>(
     flen: u64,
     field1_kind: &mut RegPassKind,
     field2_kind: &mut RegPassKind,
+    offset_from_start: Size,
 ) -> Result<(), CannotUseFpConv>
 where
     Ty: TyAbiInterface<'a, C> + Copy,
@@ -49,16 +50,16 @@ where
                 }
                 match (*field1_kind, *field2_kind) {
                     (RegPassKind::Unknown, _) => {
-                        *field1_kind = RegPassKind::Integer(Reg {
-                            kind: RegKind::Integer,
-                            size: arg_layout.size,
-                        });
+                        *field1_kind = RegPassKind::Integer {
+                            offset_from_start,
+                            ty: Reg { kind: RegKind::Integer, size: arg_layout.size },
+                        };
                     }
-                    (RegPassKind::Float(_), RegPassKind::Unknown) => {
-                        *field2_kind = RegPassKind::Integer(Reg {
-                            kind: RegKind::Integer,
-                            size: arg_layout.size,
-                        });
+                    (RegPassKind::Float { .. }, RegPassKind::Unknown) => {
+                        *field2_kind = RegPassKind::Integer {
+                            offset_from_start,
+                            ty: Reg { kind: RegKind::Integer, size: arg_layout.size },
+                        };
                     }
                     _ => return Err(CannotUseFpConv),
                 }
@@ -69,12 +70,16 @@ where
                 }
                 match (*field1_kind, *field2_kind) {
                     (RegPassKind::Unknown, _) => {
-                        *field1_kind =
-                            RegPassKind::Float(Reg { kind: RegKind::Float, size: arg_layout.size });
+                        *field1_kind = RegPassKind::Float {
+                            offset_from_start,
+                            ty: Reg { kind: RegKind::Float, size: arg_layout.size },
+                        };
                     }
                     (_, RegPassKind::Unknown) => {
-                        *field2_kind =
-                            RegPassKind::Float(Reg { kind: RegKind::Float, size: arg_layout.size });
+                        *field2_kind = RegPassKind::Float {
+                            offset_from_start,
+                            ty: Reg { kind: RegKind::Float, size: arg_layout.size },
+                        };
                     }
                     _ => return Err(CannotUseFpConv),
                 }
@@ -96,13 +101,14 @@ where
                             flen,
                             field1_kind,
                             field2_kind,
+                            offset_from_start,
                         );
                     }
                     return Err(CannotUseFpConv);
                 }
             }
             FieldsShape::Array { count, .. } => {
-                for _ in 0..count {
+                for i in 0..count {
                     let elem_layout = arg_layout.field(cx, 0);
                     should_use_fp_conv_helper(
                         cx,
@@ -111,6 +117,7 @@ where
                         flen,
                         field1_kind,
                         field2_kind,
+                        offset_from_start + elem_layout.size * i,
                     )?;
                 }
             }
@@ -121,7 +128,15 @@ where
                 }
                 for i in arg_layout.fields.index_by_increasing_offset() {
                     let field = arg_layout.field(cx, i);
-                    should_use_fp_conv_helper(cx, &field, xlen, flen, field1_kind, field2_kind)?;
+                    should_use_fp_conv_helper(
+                        cx,
+                        &field,
+                        xlen,
+                        flen,
+                        field1_kind,
+                        field2_kind,
+                        offset_from_start + arg_layout.fields.offset(i),
+                    )?;
                 }
             }
         },
@@ -140,14 +155,52 @@ where
 {
     let mut field1_kind = RegPassKind::Unknown;
     let mut field2_kind = RegPassKind::Unknown;
-    if should_use_fp_conv_helper(cx, arg, xlen, flen, &mut field1_kind, &mut field2_kind).is_err() {
+    if should_use_fp_conv_helper(
+        cx,
+        arg,
+        xlen,
+        flen,
+        &mut field1_kind,
+        &mut field2_kind,
+        Size::ZERO,
+    )
+    .is_err()
+    {
         return None;
     }
     match (field1_kind, field2_kind) {
-        (RegPassKind::Integer(l), RegPassKind::Float(r)) => Some(FloatConv::MixedPair(l, r)),
-        (RegPassKind::Float(l), RegPassKind::Integer(r)) => Some(FloatConv::MixedPair(l, r)),
-        (RegPassKind::Float(l), RegPassKind::Float(r)) => Some(FloatConv::FloatPair(l, r)),
-        (RegPassKind::Float(f), RegPassKind::Unknown) => Some(FloatConv::Float(f)),
+        (
+            RegPassKind::Integer { offset_from_start, .. }
+            | RegPassKind::Float { offset_from_start, .. },
+            _,
+        ) if offset_from_start != Size::ZERO => {
+            panic!("type {:?} has a first field with non-zero offset {offset_from_start:?}", arg.ty)
+        }
+        (
+            RegPassKind::Integer { ty: first_ty, .. },
+            RegPassKind::Float { offset_from_start, ty: second_ty },
+        ) => Some(FloatConv::MixedPair {
+            first_ty,
+            second_ty_offset_from_start: offset_from_start,
+            second_ty,
+        }),
+        (
+            RegPassKind::Float { ty: first_ty, .. },
+            RegPassKind::Integer { offset_from_start, ty: second_ty },
+        ) => Some(FloatConv::MixedPair {
+            first_ty,
+            second_ty_offset_from_start: offset_from_start,
+            second_ty,
+        }),
+        (
+            RegPassKind::Float { ty: first_ty, .. },
+            RegPassKind::Float { offset_from_start, ty: second_ty },
+        ) => Some(FloatConv::FloatPair {
+            first_ty,
+            second_ty_offset_from_start: offset_from_start,
+            second_ty,
+        }),
+        (RegPassKind::Float { ty, .. }, RegPassKind::Unknown) => Some(FloatConv::Float(ty)),
         _ => None,
     }
 }
@@ -165,11 +218,19 @@ where
             FloatConv::Float(f) => {
                 arg.cast_to(f);
             }
-            FloatConv::FloatPair(l, r) => {
-                arg.cast_to(CastTarget::pair(l, r));
+            FloatConv::FloatPair { first_ty, second_ty_offset_from_start, second_ty } => {
+                arg.cast_to(CastTarget::offset_pair(
+                    first_ty,
+                    second_ty_offset_from_start,
+                    second_ty,
+                ));
             }
-            FloatConv::MixedPair(l, r) => {
-                arg.cast_to(CastTarget::pair(l, r));
+            FloatConv::MixedPair { first_ty, second_ty_offset_from_start, second_ty } => {
+                arg.cast_to(CastTarget::offset_pair(
+                    first_ty,
+                    second_ty_offset_from_start,
+                    second_ty,
+                ));
             }
         }
         return false;
@@ -233,15 +294,27 @@ fn classify_arg<'a, Ty, C>(
                 arg.cast_to(f);
                 return;
             }
-            Some(FloatConv::FloatPair(l, r)) if *avail_fprs >= 2 => {
+            Some(FloatConv::FloatPair { first_ty, second_ty_offset_from_start, second_ty })
+                if *avail_fprs >= 2 =>
+            {
                 *avail_fprs -= 2;
-                arg.cast_to(CastTarget::pair(l, r));
+                arg.cast_to(CastTarget::offset_pair(
+                    first_ty,
+                    second_ty_offset_from_start,
+                    second_ty,
+                ));
                 return;
             }
-            Some(FloatConv::MixedPair(l, r)) if *avail_fprs >= 1 && *avail_gprs >= 1 => {
+            Some(FloatConv::MixedPair { first_ty, second_ty_offset_from_start, second_ty })
+                if *avail_fprs >= 1 && *avail_gprs >= 1 =>
+            {
                 *avail_gprs -= 1;
                 *avail_fprs -= 1;
-                arg.cast_to(CastTarget::pair(l, r));
+                arg.cast_to(CastTarget::offset_pair(
+                    first_ty,
+                    second_ty_offset_from_start,
+                    second_ty,
+                ));
                 return;
             }
             _ => (),
diff --git a/tests/assembly/loongarch-float-struct-abi.rs b/tests/assembly/loongarch-float-struct-abi.rs
new file mode 100644
index 0000000..4991004
--- /dev/null
+++ b/tests/assembly/loongarch-float-struct-abi.rs
@@ -0,0 +1,134 @@
+//@ add-core-stubs
+//@ assembly-output: emit-asm
+//@ compile-flags: -Copt-level=3 --target loongarch64-unknown-linux-gnu
+//@ needs-llvm-components: loongarch
+
+#![feature(no_core, lang_items)]
+#![no_std]
+#![no_core]
+#![crate_type = "lib"]
+
+extern crate minicore;
+use minicore::*;
+
+#[repr(C, align(64))]
+struct Aligned(f64);
+
+#[repr(C)]
+struct Padded(u8, Aligned);
+
+#[repr(C, packed)]
+struct Packed(u8, f32);
+
+impl Copy for Aligned {}
+impl Copy for Padded {}
+impl Copy for Packed {}
+
+extern "C" {
+    fn take_padded(x: Padded);
+    fn get_padded() -> Padded;
+    fn take_packed(x: Packed);
+    fn get_packed() -> Packed;
+}
+
+// CHECK-LABEL: pass_padded
+#[unsafe(no_mangle)]
+extern "C" fn pass_padded(out: &mut Padded, x: Padded) {
+    // CHECK: st.b $a1, $a0, 0
+    // CHECK-NEXT: fst.d $fa0, $a0, 64
+    // CHECK-NEXT: ret
+    *out = x;
+}
+
+// CHECK-LABEL: ret_padded
+#[unsafe(no_mangle)]
+extern "C" fn ret_padded(x: &Padded) -> Padded {
+    // CHECK: fld.d $fa0, $a0, 64
+    // CHECK-NEXT: ld.b $a0, $a0, 0
+    // CHECK-NEXT: ret
+    *x
+}
+
+#[unsafe(no_mangle)]
+extern "C" fn call_padded(x: &Padded) {
+    // CHECK: fld.d $fa0, $a0, 64
+    // CHECK-NEXT: ld.b $a0, $a0, 0
+    // CHECK-NEXT: pcaddu18i $t8, %call36(take_padded)
+    // CHECK-NEXT: jr $t8
+    unsafe {
+        take_padded(*x);
+    }
+}
+
+#[unsafe(no_mangle)]
+extern "C" fn receive_padded(out: &mut Padded) {
+    // CHECK: addi.d $sp, $sp, -16
+    // CHECK-NEXT: .cfi_def_cfa_offset 16
+    // CHECK-NEXT: st.d $ra, $sp, [[#%d,RA_SPILL:]]
+    // CHECK-NEXT: st.d [[TEMP:.*]], $sp, [[#%d,TEMP_SPILL:]]
+    // CHECK-NEXT: .cfi_offset 1, [[#%d,RA_SPILL - 16]]
+    // CHECK-NEXT: .cfi_offset [[#%d,TEMP_NUM:]], [[#%d,TEMP_SPILL - 16]]
+    // CHECK-NEXT: move [[TEMP]], $a0
+    // CHECK-NEXT: pcaddu18i $ra, %call36(get_padded)
+    // CHECK-NEXT: jirl $ra, $ra, 0
+    // CHECK-NEXT: st.b $a0, [[TEMP]], 0
+    // CHECK-NEXT: fst.d $fa0, [[TEMP]], 64
+    // CHECK-NEXT: ld.d [[TEMP]], $sp, [[#%d,TEMP_SPILL]]
+    // CHECK-NEXT: ld.d $ra, $sp, [[#%d,RA_SPILL]]
+    // CHECK: addi.d $sp, $sp, 16
+    // CHECK: ret
+    unsafe {
+        *out = get_padded();
+    }
+}
+
+// CHECK-LABEL: pass_packed
+#[unsafe(no_mangle)]
+extern "C" fn pass_packed(out: &mut Packed, x: Packed) {
+    // CHECK: st.b $a1, $a0, 0
+    // CHECK-NEXT: fst.s $fa0, $a0, 1
+    // CHECK-NEXT: ret
+    *out = x;
+}
+
+// CHECK-LABEL: ret_packed
+#[unsafe(no_mangle)]
+extern "C" fn ret_packed(x: &Packed) -> Packed {
+    // CHECK: fld.s $fa0, $a0, 1
+    // CHECK-NEXT: ld.b $a0, $a0, 0
+    // CHECK-NEXT: ret
+    *x
+}
+
+#[unsafe(no_mangle)]
+extern "C" fn call_packed(x: &Packed) {
+    // CHECK: fld.s $fa0, $a0, 1
+    // CHECK-NEXT: ld.b $a0, $a0, 0
+    // CHECK-NEXT: pcaddu18i $t8, %call36(take_packed)
+    // CHECK-NEXT: jr $t8
+    unsafe {
+        take_packed(*x);
+    }
+}
+
+#[unsafe(no_mangle)]
+extern "C" fn receive_packed(out: &mut Packed) {
+    // CHECK: addi.d $sp, $sp, -16
+    // CHECK-NEXT: .cfi_def_cfa_offset 16
+    // CHECK-NEXT: st.d $ra, $sp, [[#%d,RA_SPILL:]]
+    // CHECK-NEXT: st.d [[TEMP:.*]], $sp, [[#%d,TEMP_SPILL:]]
+    // CHECK-NEXT: .cfi_offset 1, [[#%d,RA_SPILL - 16]]
+    // CHECK-NEXT: .cfi_offset [[#%d,TEMP_NUM:]], [[#%d,TEMP_SPILL - 16]]
+    // CHECK-NEXT: move [[TEMP]], $a0
+    // CHECK-NEXT: pcaddu18i $ra, %call36(get_packed)
+    // CHECK-NEXT: jirl $ra, $ra, 0
+    // CHECK-NEXT: st.b $a0, [[TEMP]], 0
+    // CHECK-NEXT: fst.s $fa0, [[TEMP]], 1
+    // CHECK-NEXT: ld.d [[TEMP]], $sp, [[#%d,TEMP_SPILL]]
+    // CHECK-NEXT: ld.d $ra, $sp, [[#%d,RA_SPILL]]
+    // CHECK: addi.d $sp, $sp, 16
+    // CHECK: ret
+    unsafe {
+        *out = get_packed();
+    }
+}
diff --git a/tests/codegen/cast-target-abi.rs b/tests/codegen/cast-target-abi.rs
index cbd49e2..af3fc5b 100644
--- a/tests/codegen/cast-target-abi.rs
+++ b/tests/codegen/cast-target-abi.rs
@@ -171,7 +171,15 @@ pub extern "C" fn receives_doubledouble(x: DoubleDouble) {
 
     // CHECK: [[RUST_ALLOCA:%.+]] = alloca [16 x i8], align [[RUST_ALIGN:8]]
 
-    // CHECK: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // aarch64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = extractvalue [[ABI_TYPE]] [[ABI_VALUE:%.+]], 0
+    // loongarch64: [[ABI_VALUE_1:%.+]] = extractvalue [[ABI_TYPE]] [[ABI_VALUE:%.+]], 1
+    // loongarch64: store double [[ABI_VALUE_0]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: store double [[ABI_VALUE_1]], ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
+    // powerpc64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // sparc64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // x86_64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
     // CHECK: call void @llvm.memcpy.{{.+}}(ptr align [[RUST_ALIGN]] [[RUST_ALLOCA]], ptr align [[ABI_ALIGN]] [[ABI_ALLOCA]], i64 16, i1 false)
 }
@@ -190,7 +198,11 @@ pub extern "C" fn returns_doubledouble() -> DoubleDouble {
     // x86_64:      [[ABI_ALLOCA:%.+]] = alloca [16 x i8], align [[ABI_ALIGN:8]]
 
     // aarch64:     [[ABI_VALUE:%.+]] = load [[ABI_TYPE:\[2 x double\]]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = load double, ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: [[ABI_VALUE_1:%.+]] = load double, ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_2:%.+]] = insertvalue [[ABI_TYPE:{ double, double }]] poison, double [[ABI_VALUE_0]], 0
+    // loongarch64: [[ABI_VALUE:%.+]] = insertvalue { double, double } [[ABI_VALUE_2]], double [[ABI_VALUE_1]], 1
     // sparc64:     [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
     // x86_64:      [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
@@ -269,7 +281,11 @@ pub extern "C" fn receives_doublefloat(x: DoubleFloat) {
     // x86_64:      [[RUST_ALLOCA:%.+]] = alloca [16 x i8], align [[RUST_ALIGN:8]]
 
     // aarch64:     store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = extractvalue { double, float } [[ABI_VALUE]], 0
+    // loongarch64: [[ABI_VALUE_1:%.+]] = extractvalue { double, float } [[ABI_VALUE]], 1
+    // loongarch64: store double [[ABI_VALUE_0]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: store float [[ABI_VALUE_1]], ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
     // powerpc64:   store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
     // x86_64:      store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
@@ -297,7 +313,11 @@ pub extern "C" fn returns_doublefloat() -> DoubleFloat {
     // x86_64:      [[ABI_ALLOCA:%.+]] = alloca [16 x i8], align [[ABI_ALIGN:8]]
 
     // aarch64:     [[ABI_VALUE:%.+]] = load [[ABI_TYPE:\[2 x i64\]]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, float }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = load double, ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: [[ABI_VALUE_1:%.+]] = load float, ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_2:%.+]] = insertvalue [[ABI_TYPE:{ double, float }]] poison, double [[ABI_VALUE_0]], 0
+    // loongarch64: [[ABI_VALUE:%.+]] = insertvalue { double, float } [[ABI_VALUE_2]], float [[ABI_VALUE_1]], 1
     // x86_64:      [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
     // aarch64:     ret [[ABI_TYPE]] [[ABI_VALUE]]
@@ -429,7 +449,11 @@ pub fn call_doubledouble() {
     // CHECK: call void @llvm.memcpy.{{.+}}(ptr align [[ABI_ALIGN]] [[ABI_ALLOCA]], ptr align [[RUST_ALIGN]] [[RUST_ALLOCA]], i64 16, i1 false)
 
     // aarch64:     [[ABI_VALUE:%.+]] = load [[ABI_TYPE:\[2 x double\]]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = load double, ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: [[ABI_VALUE_1:%.+]] = load double, ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_2:%.+]] = insertvalue [[ABI_TYPE:{ double, double }]] poison, double [[ABI_VALUE_0]], 0
+    // loongarch64: [[ABI_VALUE:%.+]] = insertvalue { double, double } [[ABI_VALUE_2]], double [[ABI_VALUE_1]], 1
     // powerpc64:   [[ABI_VALUE:%.+]] = load [[ABI_TYPE:\[2 x i64\]]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
     // sparc64:     [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
     // x86_64:      [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
@@ -465,7 +489,11 @@ pub fn return_doubledouble() -> DoubleDouble {
     // x86_64:      [[ABI_VALUE:%.+]] = call [[ABI_TYPE:{ double, double }]] @returns_doubledouble()
 
     // aarch64:     store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = extractvalue { double, double } [[ABI_VALUE]], 0
+    // loongarch64: [[ABI_VALUE_1:%.+]] = extractvalue { double, double } [[ABI_VALUE]], 1
+    // loongarch64: store double [[ABI_VALUE_0]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: store double [[ABI_VALUE_1]], ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
     // sparc64:     store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
     // x86_64:      store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
@@ -500,7 +528,11 @@ pub fn call_doublefloat() {
     // x86_64:      call void @llvm.memcpy.{{.+}}(ptr align [[ABI_ALIGN]] [[ABI_ALLOCA]], ptr align [[RUST_ALIGN]] [[RUST_ALLOCA]], i64 16, i1 false)
 
     // aarch64:     [[ABI_VALUE:%.+]] = load [[ABI_TYPE:\[2 x i64\]]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, float }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = load double, ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: [[ABI_VALUE_1:%.+]] = load float, ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_2:%.+]] = insertvalue [[ABI_TYPE:{ double, float }]] poison, double [[ABI_VALUE_0]], 0
+    // loongarch64: [[ABI_VALUE:%.+]] = insertvalue { double, float } [[ABI_VALUE_2]], float [[ABI_VALUE_1]], 1
     // powerpc64:   [[ABI_VALUE:%.+]] = load [[ABI_TYPE:\[2 x i64\]]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
     // x86_64:      [[ABI_VALUE:%.+]] = load [[ABI_TYPE:{ double, double }]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
@@ -540,7 +572,11 @@ pub fn return_doublefloat() -> DoubleFloat {
     // x86_64:      [[ABI_VALUE:%.+]] = call [[ABI_TYPE:{ double, double }]] @returns_doublefloat()
 
     // aarch64:     store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
-    // loongarch64: store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_VALUE_0:%.+]] = extractvalue { double, float } [[ABI_VALUE]], 0
+    // loongarch64: [[ABI_VALUE_1:%.+]] = extractvalue { double, float } [[ABI_VALUE]], 1
+    // loongarch64: store double [[ABI_VALUE_0]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
+    // loongarch64: [[ABI_ALLOCA_1:%.+]] = getelementptr inbounds i8, ptr [[ABI_ALLOCA]], i64 8
+    // loongarch64: store float [[ABI_VALUE_1]], ptr [[ABI_ALLOCA_1]], align [[ABI_ALIGN]]
     // x86_64:      store [[ABI_TYPE]] [[ABI_VALUE]], ptr [[ABI_ALLOCA]], align [[ABI_ALIGN]]
 
     // aarch64:     call void @llvm.memcpy.{{.+}}(ptr align [[RUST_ALIGN]] [[RUST_ALLOCA]], ptr align [[ABI_ALIGN]] [[ABI_ALLOCA]], i64 16, i1 false)
diff --git a/tests/codegen/loongarch-abi/cast-local-large-enough.rs b/tests/codegen/loongarch-abi/cast-local-large-enough.rs
new file mode 100644
index 0000000..e5a0e4c
--- /dev/null
+++ b/tests/codegen/loongarch-abi/cast-local-large-enough.rs
@@ -0,0 +1,44 @@
+//@ add-core-stubs
+//@ compile-flags: -Copt-level=0 -Cdebuginfo=0 --target loongarch64-unknown-linux-gnu
+//@ needs-llvm-components: loongarch
+
+#![feature(no_core, lang_items)]
+#![no_std]
+#![no_core]
+#![crate_type = "lib"]
+
+extern crate minicore;
+use minicore::*;
+
+#[repr(C, align(64))]
+struct Aligned(f64);
+
+#[repr(C, align(64))]
+struct AlignedPair(f32, f64);
+
+impl Copy for Aligned {}
+impl Copy for AlignedPair {}
+
+// CHECK-LABEL: define double @read_aligned
+#[unsafe(no_mangle)]
+pub extern "C" fn read_aligned(x: &Aligned) -> Aligned {
+    // CHECK: %[[TEMP:.*]] = alloca [64 x i8], align 64
+    // CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 64 %[[TEMP]], ptr align 64 %[[PTR:.*]], i64 64, i1 false)
+    // CHECK-NEXT: %[[RES:.*]] = load double, ptr %[[TEMP]], align 64
+    // CHECK-NEXT: ret double %[[RES]]
+    *x
+}
+
+// CHECK-LABEL: define { float, double } @read_aligned_pair
+#[unsafe(no_mangle)]
+pub extern "C" fn read_aligned_pair(x: &AlignedPair) -> AlignedPair {
+    // CHECK: %[[TEMP:.*]] = alloca [64 x i8], align 64
+    // CHECK-NEXT: call void @llvm.memcpy.p0.p0.i64(ptr align 64 %[[TEMP]], ptr align 64 %[[PTR:.*]], i64 64, i1 false)
+    // CHECK-NEXT: %[[FIRST:.*]] = load float, ptr %[[TEMP]], align 64
+    // CHECK-NEXT: %[[SECOND_PTR:.*]] = getelementptr inbounds i8, ptr %[[TEMP]], i64 8
+    // CHECK-NEXT: %[[SECOND:.*]] = load double, ptr %[[SECOND_PTR]], align 8
+    // CHECK-NEXT: %[[RES1:.*]] = insertvalue { float, double } poison, float %[[FIRST]], 0
+    // CHECK-NEXT: %[[RES2:.*]] = insertvalue { float, double } %[[RES1]], double %[[SECOND]], 1
+    // CHECK-NEXT: ret { float, double } %[[RES2]]
+    *x
+}