SIGBUS on AArch64 when `Proc` returns large extern struct by value

[HertzDevil]

This example from spec/compiler/codegen/extern_spec.cr, which currently runs only on x86-64 targets, crashes on AArch64 (both macOS and Alpine):

lib LibMylib
  struct Struct
    x : Int32
    y : Int32
    z : Int32
    w : Int32
    a : Int32
  end
end

f = ->{
  s = LibMylib::Struct.new
  s.x = 1
  s.y = 2
  s.z = 3
  s.w = 4
  s.a = 5
  s
}

s = f.call
s.x &+ s.y &+ s.z &+ s.w &+ s.a

LibMylib::Struct must be passed indirectly according to the AArch64 ABI. The generated LLVM IR shows how this is done:

define i32 @__crystal_main(i32 %argc, ptr %argv) {
  ; ...
  %6 = load %"->", ptr %f, align 8
  %7 = extractvalue %"->" %6, 0
  %8 = extractvalue %"->" %6, 1
  %9 = icmp eq ptr %8, null
  br i1 %9, label %ctx_is_null, label %ctx_is_not_null

ctx_is_null:                                      ; preds = %entry
  call void %7(ptr %1)
  br label %exit

ctx_is_not_null:                                  ; preds = %entry
  %10 = call %"struct.LibMylib::Struct" %7(ptr %8)
  store %"struct.LibMylib::Struct" %10, ptr %2, align 4
  br label %exit

exit:                                             ; preds = %ctx_is_not_null, %ctx_is_null
  %11 = phi ptr [ %1, %ctx_is_null ], [ %2, %ctx_is_not_null ]
  %12 = load %"struct.LibMylib::Struct", ptr %11, align 4
  store %"struct.LibMylib::Struct" %12, ptr %s, align 4
  ; ...
}

define void @"~procProc(LibMylib::Struct)@test.cr:11"(ptr sret(%"struct.LibMylib::Struct") %0) #0 {
alloca:
  %s = alloca %"struct.LibMylib::Struct", align 8
  %1 = alloca %"struct.LibMylib::Struct", align 8
  br label %entry

entry:                                            ; preds = %alloca
  %2 = call %"struct.LibMylib::Struct" @"*struct.LibMylib::Struct::new:struct.LibMylib::Struct"()
  store %"struct.LibMylib::Struct" %2, ptr %1, align 4
  %3 = load %"struct.LibMylib::Struct", ptr %1, align 4
  store %"struct.LibMylib::Struct" %3, ptr %s, align 4
  %4 = getelementptr inbounds %"struct.LibMylib::Struct", ptr %s, i32 0, i32 0
  store i32 1, ptr %4, align 4
  %5 = getelementptr inbounds %"struct.LibMylib::Struct", ptr %s, i32 0, i32 1
  store i32 2, ptr %5, align 4
  %6 = getelementptr inbounds %"struct.LibMylib::Struct", ptr %s, i32 0, i32 2
  store i32 3, ptr %6, align 4
  %7 = getelementptr inbounds %"struct.LibMylib::Struct", ptr %s, i32 0, i32 3
  store i32 4, ptr %7, align 4
  %8 = getelementptr inbounds %"struct.LibMylib::Struct", ptr %s, i32 0, i32 4
  store i32 5, ptr %8, align 4
  %9 = load %"struct.LibMylib::Struct", ptr %s, align 4
  store %"struct.LibMylib::Struct" %9, ptr %0, align 4
  ret void
}

Apparently, the store %"struct.LibMylib::Struct" %9, ptr %0 instruction is reading %0 from the wrong register, but everything works if the call is adjusted like this:

ctx_is_null:                                      ; preds = %entry
  call void %7(ptr sret(%"struct.LibMylib::Struct") %1)
  br label %exit

That is, the sret at the call argument should match the sret at the function parameter.