[PATCH] Fix constant folding and poor optimization in byte swappingcode

From: Trent Piepho
Date: Fri Feb 16 2007 - 19:19:37 EST


Constant folding does not work for the swabXX() byte swapping functions,
and the C versions optimize poorly.

Attempting to initialize a global variable to swab16(0x1234) or put
something like "case swab32(42):" in a switch statement will not compile.
It can work, swab.h just isn't doing it correctly. This patch fixes
that.

Contrary to the comment in asm-i386/byteorder.h, gcc does not recognize
the "C" version of swab16 and turn it into efficient code. gcc can do
this, just not with the current code. The simple function:

u16 foo(u16 x) { return swab16(x); }

Would compile to:
movzwl %ax, %eax
movl %eax, %edx
shrl $8, %eax
sall $8, %edx
orl %eax, %edx

With this patch, it will compile to:
rolw $8, %ax

I also attempted to document the maze different macros/inline functions
that are used to create the final product.Fix constant folding and poor optimization in byte swapping code
From: Trent Piepho <xyzzy@xxxxxxxxxxxxx>

Constant folding does not work for the swabXX() byte swapping functions,
and the C versions optimize poorly.

Attempting to initialize a global variable to swab16(0x1234) or put
something like "case swab32(42):" in a switch statement will not compile.
It can work, swab.h just isn't doing it correctly. This patch fixes
that.

Contrary to the comment in asm-i386/byteorder.h, gcc does not recognize
the "C" version of swab16 and turn it into efficient code. gcc can do
this, just not with the current code. The simple function:

u16 foo(u16 x) { return swab16(x); }

Would compile to:
movzwl %ax, %eax
movl %eax, %edx
shrl $8, %eax
sall $8, %edx
orl %eax, %edx

With this patch, it will compile to:
rolw $8, %ax

I also attempted to document the maze different macros/inline functions
that are used to create the final product.

Signed-off-by: Trent Piepho <xyzzy@xxxxxxxxxxxxx>
CC: Francois-Rene Rideau <fare@xxxxxxxxx>
---

swab.h | 110 +++++++++++++++++++++++++++++++++++++++++------------------------
1 files changed, 70 insertions(+), 40 deletions(-)

diff -r 2a44233ff44e -r dff6a3ea12f1 include/linux/byteorder/swab.h
--- a/include/linux/byteorder/swab.h Tue Feb 13 22:02:32 2007 -0800
+++ b/include/linux/byteorder/swab.h Fri Feb 16 15:53:06 2007 -0800
@@ -10,6 +10,10 @@
* separated swab functions from cpu_to_XX,
* to clean up support for bizarre-endian architectures.
*
+ * Trent Piepho <xyzzy@xxxxxxxxxxxxx> 2007114
+ * make constant-folding work, provide C versions that
+ * gcc can optimize better, explain different versions
+ *
* See asm-i386/byteorder.h and suches for examples of how to provide
* architecture-dependent optimized versions
*
@@ -17,40 +21,66 @@

#include <linux/compiler.h>

+/* Functions/macros defined, there are a lot:
+ *
+ * ___swabXX
+ * Generic C versions of the swab functions.
+ *
+ * ___constant_swabXX
+ * C versions that gcc can fold into a compile-time constant when
+ * the argument is a compile-time constant.
+ *
+ * __arch__swabXX[sp]?
+ * Architecture optimized versions of all the swab functions
+ * (including the s and p versions). These can be defined in
+ * asm-arch/byteorder.h. Any which are not, are defined here.
+ * __arch__swabXXs() is defined in terms of __arch__swabXXp(), which
+ * is defined in terms of __arch__swabXX(), which is in turn defined
+ * in terms of ___swabXX(x).
+ * These must be macros. They may be unsafe for arguments with
+ * side-effects.
+ *
+ * __fswabXX
+ * Inline function versions of the __arch__ macros. These _are_ safe
+ * if the arguments have side-effects. Note there are no s and p
+ * versions of these.
+ *
+ * __swabXX[sb]
+ * There are the ones you should actually use. The __swabXX versions
+ * will be a constant given a constant argument and use the arch
+ * specific code (if any) for non-constant arguments. The s and p
+ * versions always use the arch specific code (constant folding
+ * doesn't apply). They are safe to use with arguments with
+ * side-effects.
+ *
+ * swabXX[sb]
+ * Nicknames for __swabXX[sb] to use in the kernel.
+ */
+
/* casts are necessary for constants, because we never know how for sure
* how U/UL/ULL map to __u16, __u32, __u64. At least not in a portable way.
*/
-#define ___swab16(x) \
-({ \
- __u16 __x = (x); \
- ((__u16)( \
- (((__u16)(__x) & (__u16)0x00ffU) << 8) | \
- (((__u16)(__x) & (__u16)0xff00U) >> 8) )); \
-})
-
-#define ___swab32(x) \
-({ \
- __u32 __x = (x); \
- ((__u32)( \
- (((__u32)(__x) & (__u32)0x000000ffUL) << 24) | \
- (((__u32)(__x) & (__u32)0x0000ff00UL) << 8) | \
- (((__u32)(__x) & (__u32)0x00ff0000UL) >> 8) | \
- (((__u32)(__x) & (__u32)0xff000000UL) >> 24) )); \
-})
-
-#define ___swab64(x) \
-({ \
- __u64 __x = (x); \
- ((__u64)( \
- (__u64)(((__u64)(__x) & (__u64)0x00000000000000ffULL) << 56) | \
- (__u64)(((__u64)(__x) & (__u64)0x000000000000ff00ULL) << 40) | \
- (__u64)(((__u64)(__x) & (__u64)0x0000000000ff0000ULL) << 24) | \
- (__u64)(((__u64)(__x) & (__u64)0x00000000ff000000ULL) << 8) | \
- (__u64)(((__u64)(__x) & (__u64)0x000000ff00000000ULL) >> 8) | \
- (__u64)(((__u64)(__x) & (__u64)0x0000ff0000000000ULL) >> 24) | \
- (__u64)(((__u64)(__x) & (__u64)0x00ff000000000000ULL) >> 40) | \
- (__u64)(((__u64)(__x) & (__u64)0xff00000000000000ULL) >> 56) )); \
-})
+
+static __inline__ __attribute_const__ __u16 ___swab16(__u16 x)
+{
+ return x<<8 | x>>8;
+}
+static __inline__ __attribute_const__ __u32 ___swab32(__u32 x)
+{
+ return x<<24 | x>>24 |
+ (x & (__u32)0x0000ff00UL)<<8 |
+ (x & (__u32)0x00ff0000UL)>>8;
+}
+static __inline__ __attribute_const__ __u64 ___swab64(__u64 x)
+{
+ return x<<56 | x>>56 |
+ (x & (__u64)0x000000000000ff00ULL)<<40 |
+ (x & (__u64)0x0000000000ff0000ULL)<<24 |
+ (x & (__u64)0x00000000ff000000ULL)<< 8 |
+ (x & (__u64)0x000000ff00000000ULL)>> 8 |
+ (x & (__u64)0x0000ff0000000000ULL)>>24 |
+ (x & (__u64)0x00ff000000000000ULL)>>40;
+}

#define ___constant_swab16(x) \
((__u16)( \
@@ -77,13 +107,13 @@
* provide defaults when no architecture-specific optimization is detected
*/
#ifndef __arch__swab16
-# define __arch__swab16(x) ({ __u16 __tmp = (x) ; ___swab16(__tmp); })
+# define __arch__swab16(x) ___swab16(x)
#endif
#ifndef __arch__swab32
-# define __arch__swab32(x) ({ __u32 __tmp = (x) ; ___swab32(__tmp); })
+# define __arch__swab32(x) ___swab32(x)
#endif
#ifndef __arch__swab64
-# define __arch__swab64(x) ({ __u64 __tmp = (x) ; ___swab64(__tmp); })
+# define __arch__swab64(x) ___swab64(x)
#endif

#ifndef __arch__swab16p
@@ -97,13 +127,13 @@
#endif

#ifndef __arch__swab16s
-# define __arch__swab16s(x) do { *(x) = __arch__swab16p((x)); } while (0)
+# define __arch__swab16s(x) ((void)(*(x) = __arch__swab16p(x)))
#endif
#ifndef __arch__swab32s
-# define __arch__swab32s(x) do { *(x) = __arch__swab32p((x)); } while (0)
+# define __arch__swab32s(x) ((void)(*(x) = __arch__swab32p(x)))
#endif
#ifndef __arch__swab64s
-# define __arch__swab64s(x) do { *(x) = __arch__swab64p((x)); } while (0)
+# define __arch__swab64s(x) ((void)(*(x) = __arch__swab64p(x)))
#endif


@@ -113,15 +143,15 @@
#if defined(__GNUC__) && defined(__OPTIMIZE__)
# define __swab16(x) \
(__builtin_constant_p((__u16)(x)) ? \
- ___swab16((x)) : \
+ ___constant_swab16((x)) : \
__fswab16((x)))
# define __swab32(x) \
(__builtin_constant_p((__u32)(x)) ? \
- ___swab32((x)) : \
+ ___constant_swab32((x)) : \
__fswab32((x)))
# define __swab64(x) \
(__builtin_constant_p((__u64)(x)) ? \
- ___swab64((x)) : \
+ ___constant_swab64((x)) : \
__fswab64((x)))
#else
# define __swab16(x) __fswab16(x)