diff --git a/llvm/test/CodeGen/ARM/fp16-return-pr60510.ll b/llvm/test/CodeGen/ARM/fp16-return-pr60510.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/fp16-return-pr60510.ll
@@ -0,0 +1,144 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+
+; No FP16/BF16
+; RUN: llc -mtriple=arm-none-eabi -float-abi=soft -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-SOFT
+; R UN: llc -mtriple=thumb-none-eabi -float-abi=soft -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-SOFT
+; RUN: llc -mtriple=arm-none-eabi -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-SOFT
+; R UN: llc -mtriple=thumb-none-eabi -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-SOFT
+; RUN: llc -mtriple=arm-none-eabihf -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-HARD
+; R UN: llc -mtriple=thumb-none-eabihf -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-HARD
+
+; FP16/BF16
+; RUN: llc -mtriple=arm-none-eabi -float-abi=soft -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFT
+; R UN: llc -mtriple=thumb-none-eabi -float-abi=soft -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFT
+; RUN: llc -mtriple=arm-none-eabi -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFT
+; R UN: llc -mtriple=thumb-none-eabi -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFT
+; RUN: llc -mtriple=arm-none-eabihf -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-HARD
+; R UN: llc -mtriple=thumb-none-eabihf -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-HARD
+
+
+; PR60510 showed a bug where the return from `*_inner` was getting "lost" by an
+; optimisation, and a garbage value was being left in `s0`.
+
+declare dso_local float @other(float) nounwind
+
+declare dso_local bfloat @fp16_inner() nounwind
+
+define half @fp16_out_call(float %arg) nounwind {
+; NO-FP16-SOFT-LABEL: fp16_out_call:
+; NO-FP16-SOFT:       @ %bb.0:
+; NO-FP16-SOFT-NEXT:    .save {r4, r5, r11, lr}
+; NO-FP16-SOFT-NEXT:    push {r4, r5, r11, lr}
+; NO-FP16-SOFT-NEXT:    mov r4, r0
+; NO-FP16-SOFT-NEXT:    bl fp16_inner
+; NO-FP16-SOFT-NEXT:    mov r5, r0
+; NO-FP16-SOFT-NEXT:    mov r0, r4
+; NO-FP16-SOFT-NEXT:    bl other
+; NO-FP16-SOFT-NEXT:    mov r0, r5
+; NO-FP16-SOFT-NEXT:    pop {r4, r5, r11, pc}
+;
+; NO-FP16-HARD-LABEL: fp16_out_call:
+; NO-FP16-HARD:       @ %bb.0:
+; NO-FP16-HARD-NEXT:    .save {r11, lr}
+; NO-FP16-HARD-NEXT:    push {r11, lr}
+; NO-FP16-HARD-NEXT:    .vsave {d8, d9}
+; NO-FP16-HARD-NEXT:    vpush {d8, d9}
+; NO-FP16-HARD-NEXT:    vmov.f32 s16, s0
+; NO-FP16-HARD-NEXT:    bl fp16_inner
+; NO-FP16-HARD-NEXT:    vmov.f32 s18, s0
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s16
+; NO-FP16-HARD-NEXT:    bl other
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s18
+; NO-FP16-HARD-NEXT:    vpop {d8, d9}
+; NO-FP16-HARD-NEXT:    pop {r11, pc}
+;
+; FP16-SOFT-LABEL: fp16_out_call:
+; FP16-SOFT:       @ %bb.0:
+; FP16-SOFT-NEXT:    .save {r4, r5, r11, lr}
+; FP16-SOFT-NEXT:    push {r4, r5, r11, lr}
+; FP16-SOFT-NEXT:    mov r4, r0
+; FP16-SOFT-NEXT:    bl fp16_inner
+; FP16-SOFT-NEXT:    mov r5, r0
+; FP16-SOFT-NEXT:    mov r0, r4
+; FP16-SOFT-NEXT:    bl other
+; FP16-SOFT-NEXT:    vmov.f16 s0, r5
+; FP16-SOFT-NEXT:    vmov r0, s0
+; FP16-SOFT-NEXT:    pop {r4, r5, r11, pc}
+;
+; FP16-HARD-LABEL: fp16_out_call:
+; FP16-HARD:       @ %bb.0:
+; FP16-HARD-NEXT:    .save {r11, lr}
+; FP16-HARD-NEXT:    push {r11, lr}
+; FP16-HARD-NEXT:    .vsave {d8}
+; FP16-HARD-NEXT:    vpush {d8}
+; FP16-HARD-NEXT:    vmov.f32 s16, s0
+; FP16-HARD-NEXT:    bl fp16_inner
+; FP16-HARD-NEXT:    vmov.f32 s0, s16
+; FP16-HARD-NEXT:    bl other
+; FP16-HARD-NEXT:    vpop {d8}
+; FP16-HARD-NEXT:    pop {r11, pc}
+  %call = call half @fp16_inner()
+  %call1 = call float @other(float %arg)
+  ret half %call
+}
+
+declare dso_local bfloat @bf_inner() nounwind
+
+define bfloat @bf_out_call(float %arg) nounwind {
+; NO-FP16-SOFT-LABEL: bf_out_call:
+; NO-FP16-SOFT:       @ %bb.0:
+; NO-FP16-SOFT-NEXT:    .save {r4, r5, r11, lr}
+; NO-FP16-SOFT-NEXT:    push {r4, r5, r11, lr}
+; NO-FP16-SOFT-NEXT:    mov r4, r0
+; NO-FP16-SOFT-NEXT:    bl bf_inner
+; NO-FP16-SOFT-NEXT:    mov r5, r0
+; NO-FP16-SOFT-NEXT:    mov r0, r4
+; NO-FP16-SOFT-NEXT:    bl other
+; NO-FP16-SOFT-NEXT:    mov r0, r5
+; NO-FP16-SOFT-NEXT:    pop {r4, r5, r11, pc}
+;
+; NO-FP16-HARD-LABEL: bf_out_call:
+; NO-FP16-HARD:       @ %bb.0:
+; NO-FP16-HARD-NEXT:    .save {r11, lr}
+; NO-FP16-HARD-NEXT:    push {r11, lr}
+; NO-FP16-HARD-NEXT:    .vsave {d8, d9}
+; NO-FP16-HARD-NEXT:    vpush {d8, d9}
+; NO-FP16-HARD-NEXT:    vmov.f32 s16, s0
+; NO-FP16-HARD-NEXT:    bl bf_inner
+; NO-FP16-HARD-NEXT:    vmov.f32 s18, s0
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s16
+; NO-FP16-HARD-NEXT:    bl other
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s18
+; NO-FP16-HARD-NEXT:    vpop {d8, d9}
+; NO-FP16-HARD-NEXT:    pop {r11, pc}
+;
+; FP16-SOFT-LABEL: bf_out_call:
+; FP16-SOFT:       @ %bb.0:
+; FP16-SOFT-NEXT:    .save {r4, r5, r11, lr}
+; FP16-SOFT-NEXT:    push {r4, r5, r11, lr}
+; FP16-SOFT-NEXT:    mov r4, r0
+; FP16-SOFT-NEXT:    bl bf_inner
+; FP16-SOFT-NEXT:    mov r5, r0
+; FP16-SOFT-NEXT:    mov r0, r4
+; FP16-SOFT-NEXT:    bl other
+; FP16-SOFT-NEXT:    mov r0, r5
+; FP16-SOFT-NEXT:    pop {r4, r5, r11, pc}
+;
+; FP16-HARD-LABEL: bf_out_call:
+; FP16-HARD:       @ %bb.0:
+; FP16-HARD-NEXT:    .save {r11, lr}
+; FP16-HARD-NEXT:    push {r11, lr}
+; FP16-HARD-NEXT:    .vsave {d8, d9}
+; FP16-HARD-NEXT:    vpush {d8, d9}
+; FP16-HARD-NEXT:    vmov.f32 s16, s0
+; FP16-HARD-NEXT:    bl bf_inner
+; FP16-HARD-NEXT:    vmov.f32 s18, s0
+; FP16-HARD-NEXT:    vmov.f32 s0, s16
+; FP16-HARD-NEXT:    bl other
+; FP16-HARD-NEXT:    vmov.f32 s0, s18
+; FP16-HARD-NEXT:    vpop {d8, d9}
+; FP16-HARD-NEXT:    pop {r11, pc}
+  %call = call bfloat @bf_inner()
+  %call1 = call float @other(float %arg)
+  ret bfloat %call
+}