[Bug c/16355] New: Duplicate function epilog

[Bug c/16355] New: Duplicate function epilog

[anton at samba dot org]

A 3.5 powerpc64 toolchain from cvs (20040703) was used to compile a linux
kernel. I noticed some functions had more than one function epilog. A cut down
testcase is:

int foo(a)
{
        if (a)
                return 1;
#if 1
        bar();
#endif
        return 0;
}

When the call to bar() is present we get 2 function epilogs:

.foo:
        cmpdi 7,3,0
        mflr 0
        li 3,1
        std 0,16(1)
        stdu 1,-112(1)
        beq- 7,.L6
        addi 1,1,112
        ld 0,16(1)
        mtlr 0
        blr
.L6:
        bl .bar
        nop
        addi 1,1,112
        li 3,0
        ld 0,16(1)
        mtlr 0
        blr

We only duplicate 4 instructions here, but on functions which spill a large
number of registers the duplication is worse.

-- 
           Summary: Duplicate function epilog
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: anton at samba dot org
                CC: gcc-bugs at gcc dot gnu dot org
GCC target triplet: powerpc64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16355

[Bug c/16355] Duplicate function epilog

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-07-04 07:05 -------
If you want size reduced use -Os, this is done for speed reasons since you do not have to branch back.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16355

[Bug c/16355] Duplicate function epilog

[anton at samba dot org]

------- Additional Comments From anton at samba dot org  2004-07-04 07:44 -------
If we schedule it properly the branch should be free on POWER4, 970 etc.
futex_requeue is an example of what Im worried about, the duplication is a
rather big abuser of icache:

c000000000078d74:       38 21 01 10     addi    r1,r1,272
c000000000078d78:       7f 63 db 78     mr      r3,r27
c000000000078d7c:       e8 01 00 10     ld      r0,16(r1)
c000000000078d80:       ea a1 ff a8     ld      r21,-88(r1)
c000000000078d84:       ea c1 ff b0     ld      r22,-80(r1)
c000000000078d88:       ea e1 ff b8     ld      r23,-72(r1)
c000000000078d8c:       eb 01 ff c0     ld      r24,-64(r1)
c000000000078d90:       eb 21 ff c8     ld      r25,-56(r1)
c000000000078d94:       eb 41 ff d0     ld      r26,-48(r1)
c000000000078d98:       eb 61 ff d8     ld      r27,-40(r1)
c000000000078d9c:       eb 81 ff e0     ld      r28,-32(r1)
c000000000078da0:       eb a1 ff e8     ld      r29,-24(r1)
c000000000078da4:       7c 08 03 a6     mtlr    r0
c000000000078da8:       eb c1 ff f0     ld      r30,-16(r1)
c000000000078dac:       eb e1 ff f8     ld      r31,-8(r1)
c000000000078db0:       4e 80 00 20     blr
c000000000078db4:       3b 60 ff f2     li      r27,-14
c000000000078db8:       4b ff ff 80     b       c000000000078d38
<.futex_requeue+0x12c>
c000000000078dbc:       e8 7e 80 28     ld      r3,-32728(r30)
c000000000078dc0:       38 80 01 66     li      r4,358
c000000000078dc4:       4b fd 4f 85     bl      c00000000004dd48 <.__might_sleep>
c000000000078dc8:       60 00 00 00     nop
c000000000078dcc:       7f 6a db 78     mr      r10,r27
c000000000078dd0:       81 1d 00 00     lwz     r8,0(r29)
c000000000078dd4:       4b ff ff 48     b       c000000000078d1c
<.futex_requeue+0x110>
c000000000078dd8:       48 18 22 79     bl      c0000000001fb050 <.rwsem_wake>
c000000000078ddc:       60 00 00 00     nop
c000000000078de0:       38 21 01 10     addi    r1,r1,272
c000000000078de4:       7f 63 db 78     mr      r3,r27
c000000000078de8:       e8 01 00 10     ld      r0,16(r1)
c000000000078dec:       ea a1 ff a8     ld      r21,-88(r1)
c000000000078df0:       ea c1 ff b0     ld      r22,-80(r1)
c000000000078df4:       ea e1 ff b8     ld      r23,-72(r1)
c000000000078df8:       eb 01 ff c0     ld      r24,-64(r1)
c000000000078dfc:       eb 21 ff c8     ld      r25,-56(r1)
c000000000078e00:       eb 41 ff d0     ld      r26,-48(r1)
c000000000078e04:       eb 61 ff d8     ld      r27,-40(r1)
c000000000078e08:       eb 81 ff e0     ld      r28,-32(r1)
c000000000078e0c:       eb a1 ff e8     ld      r29,-24(r1)
c000000000078e10:       7c 08 03 a6     mtlr    r0
c000000000078e14:       eb c1 ff f0     ld      r30,-16(r1)
c000000000078e18:       eb e1 ff f8     ld      r31,-8(r1)
c000000000078e1c:       4e 80 00 20     blr


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16355

[Bug c/16355] Duplicate function epilog

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-07-04 07:54 -------
Give the full example, instead of the cut down one then because right now the cut down one is faster 
with the duplication, and yes the branch can be free but usually is not.  Oh and if you do something 
werid like a lot of sib calling, you would consider that a huge hit also when in fact it can save you a 
couple of cycles instead.  Note if we dup the epilog we can schedule it with the previous instructions 
and get a little speed increase.  Also again -Os disables the duplication.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16355