[Bug regression/21198] New: Packed bitfield structure copies are very inefficient

["matt at 3am-software dot com" <gcc-bugzilla@gcc.gnu.org>]

In a digression from 2.95, gcc 4.1 copies a packed bitfield structure member by member instead of 
doing as an aggregate.

Given the program:

struct __attribute__((packed)) A { unsigned short i : 1, l : 1, j : 3, k : 11; };
struct A sA;
struct A retmeA (struct A x) { return x; }

GCC 4.1 -O3 compiles that as:

        .align 1
.globl retmeA
        .type   retmeA, @function
retmeA:
        .word 0x0
        subl2 $4,%sp
        movl %r1,%r0
        movw 4(%ap),%r2
        rotl $30,%r2,%r3
        bicl2 $-8,%r3
        rotl $31,%r2,%r4
        bicl2 $-2,%r4
        rotl $27,%r2,%r1
        bicl2 $-2048,%r1
        insv %r1,$5,$11,(%r0)
        insv %r3,$2,$3,(%r0)
        insv %r4,$1,$1,(%r0)
        insv %r2,$0,$1,(%r0)
        ret
        .size   retmeA, .-retmeA

Whereas GCC 2.95.3 compiled it to:

.globl retmeA
        .type    retmeA,@function
retmeA:
        .word 0x0
        movw 4(%ap),(%r1)
        ret

-- 
           Summary: Packed bitfield structure copies are very inefficient
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: regression
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: matt at 3am-software dot com
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: x86_64--netbsd
  GCC host triplet: x86_64--netbsd
GCC target triplet: vax--netbsdelf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21198