aarch64: Use .arch directive instead of #[target_feature]

Support outline-atomics on pre-1.61 rustc.
https://developer.arm.com/documentation/dui0774/l/armclang-Integrated-Assembler/AArch64-Target-selection-directives?lang=en

README.md

@@ -154,7 +154,7 @@ RUSTFLAGS="--cfg portable_atomic_unsafe_assume_single_core" cargo ...
   If dynamic dispatching by run-time CPU feature detection is enabled, it allows maintaining support for older CPUs while using features that are not supported on older CPUs, such as CMPXCHG16B (x86_64) and FEAT_LSE (aarch64).
 
   Note:
-  - Dynamic detection is currently only enabled in Rust 1.61+ for aarch64, in Rust 1.59+ (AVX) or 1.69+ (CMPXCHG16B) for x86_64, otherwise it works the same as when this cfg is set.
+  - Dynamic detection is currently only enabled in Rust 1.59+ for aarch64, in Rust 1.59+ (AVX) or 1.69+ (CMPXCHG16B) for x86_64, otherwise it works the same as when this cfg is set.
   - If the required target features are enabled at compile-time, the atomic operations are inlined.
   - This is compatible with no-std (as with all features except `std`).
   - Some aarch64 targets enable LLVM's `outline-atomics` target feature by default, so if you set this cfg, you may want to disable that as well. (portable-atomic's outline-atomics does not depend on the compiler-rt symbols, so even if you need to disable LLVM's outline-atomics, you may not need to disable portable-atomic's outline-atomics.)

build.rs

@@ -176,20 +176,6 @@ fn main() {
             target_feature_if("cmpxchg16b", has_cmpxchg16b, &version, Some(69), true);
         }
         "aarch64" => {
-            // aarch64_target_feature stabilized in Rust 1.61 (nightly-2022-03-16): https://github.com/rust-lang/rust/pull/90621
-            if !version.probe(61, 2022, 3, 15) {
-                if version.nightly && is_allowed_feature("aarch64_target_feature") {
-                    // The part of this feature we use has not been changed since 1.27
-                    // (https://github.com/rust-lang/rust/commit/1217d70465edb2079880347fea4baaac56895f51)
-                    // until it was stabilized in nightly-2022-03-16, so it can be safely enabled in
-                    // nightly, which is older than nightly-2022-03-16.
-                    println!("cargo:rustc-cfg=portable_atomic_unstable_aarch64_target_feature");
-                } else {
-                    // On aarch64, when aarch64_target_feature is not available, outline-atomics is also not available.
-                    println!("cargo:rustc-cfg=portable_atomic_no_outline_atomics");
-                }
-            }
-
             // aarch64 macos always support FEAT_LSE and FEAT_LSE2 because it is armv8.5-a:
             // https://github.com/llvm/llvm-project/blob/llvmorg-16.0.0/llvm/include/llvm/TargetParser/AArch64TargetParser.h#L458
             let is_macos = target_os == "macos";

src/imp/atomic128/aarch64.rs

@@ -143,6 +143,30 @@ macro_rules! debug_assert_lse {
     };
 }
 
+// https://developer.arm.com/documentation/dui0774/l/armclang-Integrated-Assembler/AArch64-Target-selection-directives?lang=en
+#[cfg(not(any(target_feature = "lse", portable_atomic_target_feature = "lse")))]
+#[cfg(any(
+    target_feature = "lse",
+    portable_atomic_target_feature = "lse",
+    not(portable_atomic_no_outline_atomics),
+))]
+macro_rules! start_lse {
+    () => {
+        ".arch armv8-a+lse"
+    };
+}
+#[cfg(any(target_feature = "lse", portable_atomic_target_feature = "lse"))]
+#[cfg(any(
+    target_feature = "lse",
+    portable_atomic_target_feature = "lse",
+    not(portable_atomic_no_outline_atomics),
+))]
+macro_rules! start_lse {
+    () => {
+        ""
+    };
+}
+
 #[cfg(target_pointer_width = "32")]
 macro_rules! ptr_modifier {
     () => {
@@ -462,7 +486,8 @@ unsafe fn atomic_compare_exchange(
     #[cfg(not(any(target_feature = "lse", portable_atomic_target_feature = "lse")))]
     let res = {
         fn_alias! {
-            #[target_feature(enable = "lse")]
+            // inline(never) is not strictly necessary, but is used for clarity.
+            #[inline(never)]
             unsafe fn(dst: *mut u128, old: u128, new: u128) -> u128;
             atomic_compare_exchange_casp_relaxed
                 = _atomic_compare_exchange_casp(Ordering::Relaxed);
@@ -570,10 +595,11 @@ unsafe fn atomic_compare_exchange(
     portable_atomic_target_feature = "lse",
     not(portable_atomic_no_outline_atomics),
 ))]
-#[cfg_attr(
-    not(any(target_feature = "lse", portable_atomic_target_feature = "lse")),
-    target_feature(enable = "lse")
-)]
+// Note: If FEAT_LSE is not available at compile-time, we must guarantee that
+// the eventual caller will not be inlined. Otherwise, optimizations that reorder
+// its instructions across the if condition might introduce undefined behavior.
+// However, our code uses the ifunc helper macro, so we usually don't have to
+// worry about this.
 #[inline]
 unsafe fn _atomic_compare_exchange_casp(
     dst: *mut u128,
@@ -598,6 +624,7 @@ unsafe fn _atomic_compare_exchange_casp(
         macro_rules! cmpxchg {
             ($acquire:tt, $release:tt, $fence:tt) => {
                 asm!(
+                    start_lse!(),
                     concat!("casp", $acquire, $release, " x6, x7, x4, x5, [{dst", ptr_modifier!(), "}]"),
                     $fence,
                     dst = in(reg) dst,

src/lib.rs

@@ -146,7 +146,7 @@ RUSTFLAGS="--cfg portable_atomic_unsafe_assume_single_core" cargo ...
   If dynamic dispatching by run-time CPU feature detection is enabled, it allows maintaining support for older CPUs while using features that are not supported on older CPUs, such as CMPXCHG16B (x86_64) and FEAT_LSE (aarch64).
 
   Note:
-  - Dynamic detection is currently only enabled in Rust 1.61+ for aarch64, in Rust 1.59+ (AVX) or 1.69+ (CMPXCHG16B) for x86_64, otherwise it works the same as when this cfg is set.
+  - Dynamic detection is currently only enabled in Rust 1.59+ for aarch64, in Rust 1.59+ (AVX) or 1.69+ (CMPXCHG16B) for x86_64, otherwise it works the same as when this cfg is set.
   - If the required target features are enabled at compile-time, the atomic operations are inlined.
   - This is compatible with no-std (as with all features except `std`).
   - Some aarch64 targets enable LLVM's `outline-atomics` target feature by default, so if you set this cfg, you may want to disable that as well. (portable-atomic's outline-atomics does not depend on the compiler-rt symbols, so even if you need to disable LLVM's outline-atomics, you may not need to disable portable-atomic's outline-atomics.)
@@ -247,20 +247,11 @@ RUSTFLAGS="--cfg portable_atomic_unsafe_assume_single_core" cargo ...
 // These features are already stabilized or have already been removed from compilers,
 // and can safely be enabled for old nightly as long as version detection works.
 // - cfg(target_has_atomic)
-// - #[target_feature(enable = "lse")] on AArch64
 // - #[target_feature(enable = "cmpxchg16b")] on x86_64
 // - asm! on ARM, AArch64, RISC-V, x86_64
 // - llvm_asm! on AVR (tier 3) and MSP430 (tier 3)
 // - #[instruction_set] on non-Linux pre-v6 ARM (tier 3)
 #![cfg_attr(portable_atomic_unstable_cfg_target_has_atomic, feature(cfg_target_has_atomic))]
-#![cfg_attr(
-    all(
-        target_arch = "aarch64",
-        portable_atomic_unstable_aarch64_target_feature,
-        not(portable_atomic_no_outline_atomics),
-    ),
-    feature(aarch64_target_feature)
-)]
 #![cfg_attr(
     all(
         target_arch = "x86_64",