This is related to PR42909. When a tail-call does a trivial pass-through of a
large struct, gcc generates a redundant block copy with identical source and
destination addresses. On machines like x86 it inlines that as a rep;mov, but
on others like m68k it generates a call to libc's memcpy():

> cat bug2.c
struct s1 { int x[32]; };
extern void g1(struct s1);
void f1(struct s1 s1) { g1(s1); }
> m68k-unknown-linux-gcc -O2 -fomit-frame-pointer -S bug2.c
> cat bug2.s
#NO_APP
        .file   "bug2.c"
        .text
        .align  2
        .globl  f1
        .type   f1, @function
f1:
        move.l %sp,%d0
        addq.l #4,%d0
        pea 128.w
        move.l %d0,-(%sp)
        move.l %d0,-(%sp)
        jsr memcpy
        lea (12,%sp),%sp
        jra g1
        .size   f1, .-f1
        .ident  "GCC: (GNU) 4.5.0 20100128 (experimental)"
        .section        .note.GNU-stack,"",@progbits

However, formally speaking this triggers undefined behaviour as the memcpy's
source and destination areas overlap. (Although the C standard does distinguish
between inexact and exact overlap in assignments, there's no such distinction
for library routines like memcpy.)


-- 
           Summary: invalid memcpy() in trivial tail-call with large struct
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: mikpe at it dot uu dot se
GCC target triplet: m68k-unknown-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42910

Reply via email to