Rollup merge of rust-lang#144549 - folkertdev:va-arg-arm, r=saethlin

match clang's `va_arg` assembly on arm targets

tracking issue: rust-lang#44930

For this example

```rust
#![feature(c_variadic)]

#[unsafe(no_mangle)]
unsafe extern "C" fn variadic(a: f64, mut args: ...) -> f64 {
    let b = args.arg::<f64>();
    let c = args.arg::<f64>();

    a + b + c
}
```

We currently generate (via llvm):

```asm
variadic:
    sub     sp, sp, rust-lang#12
    stmib   sp, {r2, r3}
    vmov    d0, r0, r1
    add     r0, sp, rust-lang#4
    vldr    d1, [sp, rust-lang#4]
    add     r0, r0, rust-lang#15
    bic     r0, r0, rust-lang#7
    vadd.f64        d0, d0, d1
    add     r1, r0, rust-lang#8
    str     r1, [sp]
    vldr    d1, [r0]
    vadd.f64        d0, d0, d1
    vmov    r0, r1, d0
    add     sp, sp, rust-lang#12
    bx      lr
```

LLVM is not doing a good job. In fact, it's well-known that LLVM's implementation of `va_arg` is kind of bad, and we implement it ourselves (based on clang) for many targets already. For arm,  our own `emit_ptr_va_arg` saves 3 instructions.

Next, it turns out it's important for LLVM to explicitly start and end the lifetime of the `va_list`. In rust-lang#146059 I already end the lifetime, but when looking at this again, I noticed that it is important to also start it, see https://godbolt.org/z/EGqvKTTsK: failing to explicitly start the lifetime uses an extra register.

So, the combination of `emit_ptr_va_arg` with starting/ending the lifetime makes rustc emit exactly the instructions that clang generates::

```asm
variadic:
    sub     sp, sp, rust-lang#12
    stmib   sp, {r2, r3}
    vmov    d16, r0, r1
    vldr    d17, [sp, rust-lang#4]
    vadd.f64        d16, d16, d17
    vldr    d17, [sp, rust-lang#12]
    vadd.f64        d16, d16, d17
    vmov    r0, r1, d16
    add     sp, sp, rust-lang#12
    bx      lr
```

The arguments to `emit_ptr_va_arg` are based on [the clang implementation](https://github.com/llvm/llvm-project/blob/03dc2a41f3d9a500e47b513de5c5008c06860d65/clang/lib/CodeGen/Targets/ARM.cpp#L798-L844).

r? ``@workingjubilee`` (I can re-roll if your queue is too full, but you do seem like the right person here)

try-job: armhf-gnu

Author: Zalathar

compiler/rustc_codegen_llvm/src/va_arg.rs

@@ -908,6 +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(

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -520,7 +520,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 LocalRef::Place(va_list) => {
                     bx.va_end(va_list.val.llval);
 
-                    // Explicitly end the lifetime of the `va_list`, this matters for LLVM.
+                    // Explicitly end the lifetime of the `va_list`, improves LLVM codegen.
                     bx.lifetime_end(va_list.val.llval, va_list.layout.size);
                 }
                 _ => bug!("C-variadic function must have a `VaList` place"),

compiler/rustc_codegen_ssa/src/mir/mod.rs

@@ -438,6 +438,10 @@ fn arg_local_refs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 
             if fx.fn_abi.c_variadic && arg_index == fx.fn_abi.args.len() {
                 let va_list = PlaceRef::alloca(bx, bx.layout_of(arg_ty));
+
+                // Explicitly start the lifetime of the `va_list`, improves LLVM codegen.
+                bx.lifetime_start(va_list.val.llval, va_list.layout.size);
+
                 bx.va_start(va_list.val.llval);
 
                 return LocalRef::Place(va_list);

tests/assembly-llvm/c-variadic-arm.rs

@@ -0,0 +1,26 @@
+//@ assembly-output: emit-asm
+//@ compile-flags: -Copt-level=3
+//@ only-arm
+//@ ignore-thumb
+//@ ignore-android
+#![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
+
+    // CHECK: vldr
+    // CHECK: vadd.f64
+    // CHECK: vldr
+    // CHECK: vadd.f64
+    let b = args.arg::<f64>();
+    let c = args.arg::<f64>();
+    a + b + c
+
+    // CHECK: add sp, sp
+}

tests/codegen-llvm/c-variadic-lifetime.rs

@@ -0,0 +1,21 @@
+//@ add-core-stubs
+//@ compile-flags: -Copt-level=3
+#![feature(c_variadic)]
+#![crate_type = "lib"]
+
+// Check that `%args` explicitly has its lifetime start and end. Being explicit can improve
+// instruction and register selection, see e.g. https://github.com/rust-lang/rust/pull/144549
+
+#[unsafe(no_mangle)]
+unsafe extern "C" fn variadic(a: f64, mut args: ...) -> f64 {
+    // CHECK: call void @llvm.lifetime.start.p0(i64 {{[0-9]+}}, ptr nonnull %args)
+    // CHECK: call void @llvm.va_start.p0(ptr nonnull %args)
+
+    let b = args.arg::<f64>();
+    let c = args.arg::<f64>();
+
+    a + b + c
+
+    // CHECK: call void @llvm.va_end.p0(ptr nonnull %args)
+    // CHECK: call void @llvm.lifetime.end.p0(i64 {{[0-9]+}}, ptr nonnull %args)
+}