cat: avoid unnecessary alloc on Linux

Author: oech3

src/uu/cat/src/cat.rs

@@ -488,14 +488,18 @@ fn write_fast<R: FdReadable>(handle: &mut InputHandle<R>) -> CatResult<()> {
             return Ok(());
         }
     }
+
     // If we're not on Linux or Android, or the splice() call failed,
     // fall back on slower writing.
+    write_slow(handle, stdout)
+}
+
+#[cfg_attr(any(target_os = "linux", target_os = "android"), inline(never))] // splice fast-path does not require this allocation
+#[cfg_attr(not(any(target_os = "linux", target_os = "android")), inline)]
+fn write_slow<R: FdReadable>(handle: &mut InputHandle<R>, mut stdout: io::Stdout) -> CatResult<()> {
     let mut stdout_lock = stdout.lock();
-    // stack allocation is overhead when splice succeed
-    #[cfg(any(target_os = "linux", target_os = "android"))]
-    let mut buf = vec![0; 1024 * 64];
-    #[cfg(not(any(target_os = "linux", target_os = "android")))]
     let mut buf = [0; 1024 * 64];
+
     loop {
         match handle.reader.read(&mut buf) {
             Ok(n) => {
@@ -515,8 +519,8 @@ fn write_fast<R: FdReadable>(handle: &mut InputHandle<R>) -> CatResult<()> {
     // stdout via while loop above AND there will be second splice() call
     // that will succeed, data pushed through splice will be output before
     // the data buffered in stdout.lock. Therefore additional explicit flush
-    // is required here.
-    stdout_lock.flush().inspect_err(handle_broken_pipe)?;
+    // is required here
+    stdout.flush().inspect_err(handle_broken_pipe)?;
     Ok(())
 }