[Bug rtl-optimization/15556] New: Idiom for "xchg %ah,%al" not recognized

[Bug rtl-optimization/15556] New: Idiom for "xchg %ah,%al" not recognized

[cesarb at nitnet dot com dot br]

When compiling:
#define regparm __attribute__((regparm(3)))
unsigned short regparm a(unsigned short x) { return (x << 8) | (x >> 8); }
unsigned short regparm b(unsigned short x) { return (x >> 8) | (x << 8); }

With:
gcc -W -Wall -O3 -ffast-math -fomit-frame-pointer -fexpensive-optimizations
-save-temps -march=athlon -c

The result is:
	.file	"testcase.c"
	.text
	.align 4
	.p2align 4,,15
.globl a
	.type	a, @function
a:
	movzwl	%ax, %edx
	shrw	$8, %ax
	sall	$8, %edx
	orl	%edx, %eax
	movzwl	%ax, %eax
	ret
	.size	a, .-a
	.p2align 4,,15
.globl b
	.type	b, @function
b:
	movl	%eax, %edx
	movzwl	%ax, %eax
	sall	$8, %eax
	shrw	$8, %dx
	orl	%eax, %edx
	movzwl	%dx, %eax
	ret
	.size	b, .-b
	.ident	"GCC: (GNU) 3.5.0 20040512 (experimental)"
	.section	.note.GNU-stack,"",@progbits

What I expected for both a and b is:
	xchg	%ah, %al
	ret

or maybe:
	xchg	%al, %ah
	ret

Reading specs from /usr/lib/gcc-snapshot/lib/gcc/i486-linux/3.5.0/specs
Configured with: ../src/configure -v
--enable-languages=c,c++,java,f77,objc,ada,treelang
--prefix=/usr/lib/gcc-snapshot --enable-shared --with-system-zlib --enable-nls
--enable-threads=posix --without-included-gettext --disable-werror
--enable-__cxa_atexit --enable-clocale=gnu --enable-libstdcxx-debug
--enable-java-gc=boehm --enable-java-awt=gtk i486-linux
Thread model: posix
gcc version 3.5.0 20040512 (experimental)

-- 
           Summary: Idiom for "xchg %ah,%al" not recognized
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P2
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: cesarb at nitnet dot com dot br
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15556

[Bug target/15556] Idiom for "xchg %ah,%al" not recognized

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-05-20 19:36 -------
Confirmed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
          Component|rtl-optimization            |target
     Ever Confirmed|                            |1
           Keywords|                            |missed-optimization
   Last reconfirmed|0000-00-00 00:00:00         |2004-05-20 19:36:08
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15556

[Bug target/15556] Idiom for "xchg %ah,%al" not recognized

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-05-21 01:05 -------
Notice that if you use 32bit words, it works:

------------------------------------------------------------------
#define regparm __attribute__((regparm(3)))
unsigned regparm a(unsigned x) { return (x << 24) | (x >> 8); }
unsigned regparm b(unsigned x) { return (x >> 24) | (x << 8); }
------------------------------------------------------------------

_Z1aj:
.LFB2:
 rorl $8, %eax
 ret

_Z1bj:
.LFB3:
 rorl $24, %eax
 ret

Since the optimization is done in fold-const (look for "bit_rotate"), maybe the 
bug is there, rather than in target-specific code.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15556

[Bug tree-optimization/15556] Idiom for rotations of 16bit values not recognized

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-05-21 01:50 -------
Confirmed this should be done on the tree level.  Note the pretty-print for the tree dump does not 
dump right: "  return x <<< ??? >>> 24;"

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |tree-optimization


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15556

[Bug tree-optimization/15556] Idiom for rotations of 16bit values not recognized

[bernie at develer dot com]

------- Additional Comments From bernie at develer dot com  2004-07-23 02:41 -------
The commonly used byte-swapping idiom also produces
sub-optimal code (GCC 3.4.1 on x86):

 unsigned short swab16(unsigned short x)
 {
     return (x >> 8) | (x << 8);
 }

swab16:
       movzwl  4(%esp), %edx
       movl    %edx, %eax
       shrw    $8, %ax
       sall    $8, %edx
       orl     %edx, %eax
       movzwl  %ax, %eax
       ret

Assing a few useless masks helps GCC produce
somewhat better code:


unsigned short swab16(unsigned short x)
{
  return ((x >> 8) & 0x00FF) | ((x << 8) & 0xFF00);
}

swab16:
        movl    4(%esp), %eax
        rolw    $8, %ax
        movzwl  %ax, %eax
        ret


A single xchg instruction would suffice on x86.
Both the Linux kernel and glibc are using inline
asm to workaround this missing optimization.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15556

[Bug tree-optimization/15556] Idiom for rotations of 16bit values not recognized

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-10-22 00:32 -------
These are all fixed on the mainline:
The first two examples:
a:
        rolw    $8, %ax
        movzwl  %ax, %eax
        ret
        .size   a, .-a
        .p2align 4,,15
.globl b
        .type   b, @function
b:
        rolw    $8, %ax
        movzwl  %ax, %eax
        ret
For swab16:
swab16:
        movl    4(%esp), %eax
        rolw    $8, %ax
        movzwl  %ax, %eax
        ret



-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15556