locking/atomic/x86: Rewrite x86_32 arch_atomic64_{,fetch}_{and,or,xor}() functions

Rewrite x86_32 arch_atomic64_{,fetch}_{and,or,xor}() functions to
use arch_atomic64_try_cmpxchg().  This implementation avoids one extra
trip through the CMPXCHG loop.

The value preload before the cmpxchg loop does not need to be atomic.
Use arch_atomic64_read_nonatomic(v) to load the value from atomic_t
location in a non-atomic way.

The generated code improves from:

  1917d5:	31 c9                	xor    %ecx,%ecx
  1917d7:	31 db                	xor    %ebx,%ebx
  1917d9:	89 4c 24 3c          	mov    %ecx,0x3c(%esp)
  1917dd:	8b 74 24 24          	mov    0x24(%esp),%esi
  1917e1:	89 c8                	mov    %ecx,%eax
  1917e3:	89 5c 24 34          	mov    %ebx,0x34(%esp)
  1917e7:	8b 7c 24 28          	mov    0x28(%esp),%edi
  1917eb:	21 ce                	and    %ecx,%esi
  1917ed:	89 74 24 4c          	mov    %esi,0x4c(%esp)
  1917f1:	21 df                	and    %ebx,%edi
  1917f3:	89 de                	mov    %ebx,%esi
  1917f5:	89 7c 24 50          	mov    %edi,0x50(%esp)
  1917f9:	8b 54 24 4c          	mov    0x4c(%esp),%edx
  1917fd:	8b 7c 24 2c          	mov    0x2c(%esp),%edi
  191801:	8b 4c 24 50          	mov    0x50(%esp),%ecx
  191805:	89 d3                	mov    %edx,%ebx
  191807:	89 f2                	mov    %esi,%edx
  191809:	f0 0f c7 0f          	lock cmpxchg8b (%edi)
  19180d:	89 c1                	mov    %eax,%ecx
  19180f:	8b 74 24 34          	mov    0x34(%esp),%esi
  191813:	89 d3                	mov    %edx,%ebx
  191815:	89 44 24 4c          	mov    %eax,0x4c(%esp)
  191819:	8b 44 24 3c          	mov    0x3c(%esp),%eax
  19181d:	89 df                	mov    %ebx,%edi
  19181f:	89 54 24 44          	mov    %edx,0x44(%esp)
  191823:	89 ca                	mov    %ecx,%edx
  191825:	31 de                	xor    %ebx,%esi
  191827:	31 c8                	xor    %ecx,%eax
  191829:	09 f0                	or     %esi,%eax
  19182b:	75 ac                	jne    1917d9 <...>

to:

  1912ba:	8b 06                	mov    (%esi),%eax
  1912bc:	8b 56 04             	mov    0x4(%esi),%edx
  1912bf:	89 44 24 3c          	mov    %eax,0x3c(%esp)
  1912c3:	89 c1                	mov    %eax,%ecx
  1912c5:	23 4c 24 34          	and    0x34(%esp),%ecx
  1912c9:	89 d3                	mov    %edx,%ebx
  1912cb:	23 5c 24 38          	and    0x38(%esp),%ebx
  1912cf:	89 54 24 40          	mov    %edx,0x40(%esp)
  1912d3:	89 4c 24 2c          	mov    %ecx,0x2c(%esp)
  1912d7:	89 5c 24 30          	mov    %ebx,0x30(%esp)
  1912db:	8b 5c 24 2c          	mov    0x2c(%esp),%ebx
  1912df:	8b 4c 24 30          	mov    0x30(%esp),%ecx
  1912e3:	f0 0f c7 0e          	lock cmpxchg8b (%esi)
  1912e7:	0f 85 f3 02 00 00    	jne    1915e0 <...>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: https://lore.kernel.org/r/20240410062957.322614-3-ubizjak@gmail.com

arch/x86/include/asm/atomic64_32.h

@@ -227,69 +227,62 @@ static __always_inline s64 arch_atomic64_dec_if_positive(atomic64_t *v)
 
 static __always_inline void arch_atomic64_and(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c & i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val & i));
 }
 
 static __always_inline s64 arch_atomic64_fetch_and(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c & i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val & i));
 
-	return old;
+	return val;
 }
 #define arch_atomic64_fetch_and arch_atomic64_fetch_and
 
 static __always_inline void arch_atomic64_or(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c | i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val | i));
 }
 
 static __always_inline s64 arch_atomic64_fetch_or(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c | i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val | i));
 
-	return old;
+	return val;
 }
 #define arch_atomic64_fetch_or arch_atomic64_fetch_or
 
 static __always_inline void arch_atomic64_xor(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c ^ i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val ^ i));
 }
 
 static __always_inline s64 arch_atomic64_fetch_xor(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c ^ i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val ^ i));
 
-	return old;
+	return val;
 }
 #define arch_atomic64_fetch_xor arch_atomic64_fetch_xor
 
 static __always_inline s64 arch_atomic64_fetch_add(s64 i, atomic64_t *v)
 {
-	s64 old, c = 0;
+	s64 val = arch_atomic64_read_nonatomic(v);
 
-	while ((old = arch_atomic64_cmpxchg(v, c, c + i)) != c)
-		c = old;
+	do { } while (!arch_atomic64_try_cmpxchg(v, &val, val + i));
 
-	return old;
+	return val;
 }
 #define arch_atomic64_fetch_add arch_atomic64_fetch_add