`-Zregparm` does not work correctly for LLVM builtins

[tgross35]

The following code:

use std::ffi::c_void;

unsafe extern "C" {
    fn memset(p: *mut c_void, val: i32, len: usize) -> *mut c_void;
}

#[unsafe(no_mangle)]
pub unsafe extern "C" fn entrypoint(
    len: usize, val: i32, ptr: *mut c_void
)  -> *mut c_void {
    unsafe { memset(ptr, val, len) }
}

Generates this asm:

entrypoint:
        push    esi
        sub     esp, 8
        mov     esi, ecx
        sub     esp, 4
        push    eax
        push    edx
        push    ecx
        call    memset
        add     esp, 16
        mov     eax, esi
        add     esp, 8
        pop     esi
        ret

Which isn't right: it's taking the incoming args in eax, edx, ecx (correct) and making pushing them to the stack before calling memset (incorrect). regparm should make both be passed in regs.

Generated IR:

define dso_local noundef ptr @entrypoint(i32 inreg noundef %len, i32 inreg noundef %val, ptr inreg noundef returned writeonly captures(ret: address, provenance) %ptr) unnamed_addr {
start:
  %0 = trunc i32 %val to i8
  tail call void @llvm.memset.p0.i32(ptr inreg noundef align 1 %ptr, i8 inreg noundef %0, i32 inreg noundef %len, i1 false) #2
  ret ptr %ptr
}

declare void @llvm.memset.p0.i32(ptr writeonly captures(none), i8, i32, i1 immarg) #1

If we add the following, the output looks correct:

!llvm.module.flags = !{!2}
!2 = !{i32 1, !"NumRegisterParameters", i32 3}

entrypoint:                             # @entrypoint
        push    esi
        mov     esi, eax
        mov     eax, ecx
        mov     ecx, esi
        pop     esi
        jmp     memset                          # TAILCALL

So it seems like we need to add that module flag. Clang does this as well.

Demo of everything at https://rust.godbolt.org/z/ehnTjTx9b

Tracking issue: #131749

[tgross35]

If anybody is interested in fixing this it shouldn't be too hard, the module flag just needs to be added when -Zregparm=* is passed. It also needs a test in tests/assembly-llvm.

@nikic, could this also be an LLVM bug? langref doesn't document the NumRegisterParameters module flag, and it seems like inregs on the call should allow doing this on a per-call basis.

We can pass NumRegisterParameters in any case for now though.

[winstonallo]

@rustbot claim

[winstonallo]

Hello @tgross35, I will have a look at this

[tgross35]

Awesome! I think the relevant bit should be somewhere in this function

rust/compiler/rustc_codegen_llvm/src/context.rs

Lines 162 to 166 in a153133

	pub(crate) unsafe fn create_module<'ll>(
	tcx: TyCtxt<'_>,
	llcx: &'ll llvm::Context,
	mod_name: &str,
	) -> &'ll llvm::Module {