diff --git a/compiler/rustc_codegen_llvm/src/va_arg.rs b/compiler/rustc_codegen_llvm/src/va_arg.rs
index ce079f3cb0af1..ab08125217ff5 100644
--- a/compiler/rustc_codegen_llvm/src/va_arg.rs
+++ b/compiler/rustc_codegen_llvm/src/va_arg.rs
@@ -28,9 +28,12 @@ fn round_pointer_up_to_alignment<'ll>(
     align: Align,
     ptr_ty: &'ll Type,
 ) -> &'ll Value {
-    let mut ptr_as_int = bx.ptrtoint(addr, bx.cx().type_isize());
-    ptr_as_int = round_up_to_alignment(bx, ptr_as_int, align);
-    bx.inttoptr(ptr_as_int, ptr_ty)
+    let ptr = bx.inbounds_ptradd(addr, bx.const_i32(align.bytes() as i32 - 1));
+    bx.call_intrinsic(
+        "llvm.ptrmask",
+        &[ptr_ty, bx.type_i32()],
+        &[ptr, bx.const_int(bx.isize_ty, -(align.bytes() as isize) as i64)],
+    )
 }
 
 fn emit_direct_ptr_va_arg<'ll, 'tcx>(
@@ -905,6 +908,21 @@ pub(super) fn emit_va_arg<'ll, 'tcx>(
             )
         }
         "aarch64" => emit_aapcs_va_arg(bx, addr, target_ty),
+        "arm" => {
+            // Types wider than 16 bytes are not currently supported. Clang has special logic for
+            // such types, but `VaArgSafe` is not implemented for any type that is this large.
+            assert!(bx.cx.size_of(target_ty).bytes() <= 16);
+
+            emit_ptr_va_arg(
+                bx,
+                addr,
+                target_ty,
+                PassMode::Direct,
+                SlotSize::Bytes4,
+                AllowHigherAlign::Yes,
+                ForceRightAdjust::No,
+            )
+        }
         "s390x" => emit_s390x_va_arg(bx, addr, target_ty),
         "powerpc" => emit_powerpc_va_arg(bx, addr, target_ty),
         "powerpc64" | "powerpc64le" => emit_ptr_va_arg(
diff --git a/compiler/rustc_codegen_ssa/src/mir/block.rs b/compiler/rustc_codegen_ssa/src/mir/block.rs
index bde63fd501aa2..f6d4f9a863b6d 100644
--- a/compiler/rustc_codegen_ssa/src/mir/block.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/block.rs
@@ -504,6 +504,9 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             match self.locals[mir::Local::from_usize(1 + va_list_arg_idx)] {
                 LocalRef::Place(va_list) => {
                     bx.va_end(va_list.val.llval);
+
+                    // Explicitly end the lifetime of the `va_list`, this matters for LLVM.
+                    bx.lifetime_end(va_list.val.llval, va_list.layout.size);
                 }
                 _ => bug!("C-variadic function must have a `VaList` place"),
             }
diff --git a/tests/assembly-llvm/c-variadic-arm.rs b/tests/assembly-llvm/c-variadic-arm.rs
new file mode 100644
index 0000000000000..36aecf097c222
--- /dev/null
+++ b/tests/assembly-llvm/c-variadic-arm.rs
@@ -0,0 +1,26 @@
+//@ assembly-output: emit-asm
+//@ compile-flags: -Copt-level=3
+//@ only-arm
+//@ ignore-thumb
+#![no_std]
+#![crate_type = "lib"]
+#![feature(c_variadic)]
+
+// Check that the assembly that rustc generates matches what clang emits.
+
+#[unsafe(no_mangle)]
+unsafe extern "C" fn variadic(a: f64, mut args: ...) -> f64 {
+    // CHECK-LABEL: variadic
+    // CHECK: sub sp, sp, #12
+
+    // CHECK: vldr
+    let b = args.arg::<f64>();
+    // CHECK: vldr
+    let c = args.arg::<f64>();
+
+    // CHECK: vadd.f64
+    // CHECK: vadd.f64
+    a + b + c
+
+    // CHECK: add sp, sp, #12
+}