[m68k] Miscompilation: Move immediate to status register generates incorrect assembly

[SpacePython12]

Initially ran into this error in Rust:

unsafe fn set_int_level<const LEVEL: u8>() {
    core::arch::asm!(
        "move.w #{l},%sr",
        l = const (0x2000i16 | (((LEVEL & 0x7) as i16) << 8)),
    )
}

Tried a simpler case in C and compiled using clang:
https://godbolt.org/z/eMh7qWz6G

/* Example usecase: Disabling interrupts by setting the interrupt mask bits of SR to 7. */
void disable_interrupts() {
    asm("move.w #0x2700,%sr");
}

Both backends generated IR similar to this:

define dso_local void @disable_interrupts() {
entry:
  call void asm sideeffect "move.w #0x2700,%sr", ""() #1
  ret void
}

However, something goes wrong during codegen, and somehow the move.w #imm,%sr directive turns into move.w #imm,%d0:

disable_interrupts:
 link.w	%a6, #0
 move.w	#9984, %d0
 unlk	%a6
 rts

The issue can technically be bypassed by moving the immediate value into a temporary register, and then into the status register, but it's still incorrect behavior by the codegen backend.

[llvmbot]

@llvm/issue-subscribers-backend-m68k

Author: SpacePython_ (SpacePython12)

Initially ran into this error in Rust: ```rust unsafe fn set_int_level<const LEVEL: u8>() { core::arch::asm!( "move.w #{l},%sr", l = const (0x2000i16 | (((LEVEL & 0x7) as i16) << 8)), ) } ``` Tried a simpler case in C and compiled using clang: https://godbolt.org/z/eMh7qWz6G ```c /* Example usecase: Disabling interrupts by setting the interrupt mask bits of SR to 7. */ void disable_interrupts() { asm("move.w #0x2700,%sr"); } ``` Both backends generated IR similar to this: ```ir define dso_local void @disable_interrupts() { entry: call void asm sideeffect "move.w #0x2700,%sr", ""() #1 ret void } ``` However, something goes wrong during codegen, and somehow the `move.w #imm,%sr` directive turns into `move.w #imm,%d0`: ```asm disable_interrupts: link.w %a6, #0 move.w #9984, %d0 unlk %a6 rts ``` The issue can technically be bypassed by moving the immediate value into a temporary register, and then into the status register, but it's still incorrect behavior by the codegen backend.