x86: Cleanup pthread_spin_{try}lock.S

Save a jmp on the lock path coming from an initial failure in
pthread_spin_lock.S.  This costs 4-bytes of code but since the
function still fits in the same number of 16-byte blocks (default
function alignment) it does not have affect on the total binary size
of libc.so (unchanged after this commit).

pthread_spin_trylock was using a CAS when a simple xchg works which
is often more expensive.

Full check passes on x86-64.

sysdeps/x86_64/nptl/pthread_spin_lock.S

@@ -19,18 +19,27 @@
 #include <shlib-compat.h>
 
 ENTRY(__pthread_spin_lock)
-1:	LOCK
-	decl	0(%rdi)
-	jne	2f
+	/* Always return zero.  */
 	xor	%eax, %eax
+	LOCK
+	decl	0(%rdi)
+	jne	1f
 	ret
 
 	.align	16
-2:	rep
+1:
+	/* `rep nop` == `pause`.  */
+	rep
 	nop
-	cmpl	$0, 0(%rdi)
-	jg	1b
-	jmp	2b
+	cmpl	%eax, 0(%rdi)
+	jle	1b
+	/* Just repeat the `lock decl` logic here.  The code size save
+	   of jumping back to entry doesn't change how many 16-byte
+	   chunks (default function alignment) that the code fits in.  */
+	LOCK
+	decl	0(%rdi)
+	jne	1b
+	ret
 END(__pthread_spin_lock)
 versioned_symbol (libc, __pthread_spin_lock, pthread_spin_lock, GLIBC_2_34)
 

sysdeps/x86_64/nptl/pthread_spin_trylock.S

@@ -20,13 +20,21 @@
 #include <shlib-compat.h>
 
 ENTRY(__pthread_spin_trylock)
-	movl	$1, %eax
 	xorl	%ecx, %ecx
-	lock
-	cmpxchgl %ecx, (%rdi)
+	/* xchg has implicit LOCK prefix.  */
+	xchgl	%ecx, (%rdi)
+
+	/* Branch on result.  Expectation is the use of trylock will be
+	   branching on success/failure so this branch can be used to
+	   to predict the coming branch.  It has the benefit of
+	   breaking the likely expensive memory dependency on (%rdi).  */
+	cmpl	$1, %ecx
+	jnz	1f
+	xorl	%eax, %eax
+	ret
+1:
 	movl	$EBUSY, %eax
-	cmovel	%ecx, %eax
-	retq
+	ret
 END(__pthread_spin_trylock)
 versioned_symbol (libc, __pthread_spin_trylock, pthread_spin_trylock,
 		  GLIBC_2_34)