[Bug c/27055] New: Structures are copied byte by byte into function arguments

[Bug c/27055] New: Structures are copied byte by byte into function arguments

[guillaume dot melquiond at ens-lyon dot fr]

I tried the following testcase with various GCC versions available as Debian
packages. With 3.3.5, the copy from *a to the stack frame of g is done
word-by-word with rep movsl. With 3.4.4, it is done by memcpy. Both previous
methods are fine. With 3.4.6, the copy is done byte-by-byte without string
opcodes. With 4.0.3 and 4.1.0, it is done byte-by-byte and out-of-line: there
are two jumps for each copied byte! So argument copy got broken for x86 during
GCC 3.4 cycle and it did not get any better with GCC 4.

typedef struct A { int a[1000]; } A;
void g(A);
void f(A *a) { g(*a); }

Assembly output with GCC4 and -O3 optimization:

        pushl   %ebp
        xorl    %edx, %edx
        movl    %esp, %ebp
        subl    $4008, %esp
        movl    8(%ebp), %ecx
.L3:
        cmpl    $4000, %edx
        jb      .L2
        call    g
        leave
        ret
        .p2align 4,,7
.L2:
        movzbl  (%ecx,%edx), %eax
        movb    %al, (%esp,%edx)
        incl    %edx
        .p2align 4,,3
        jmp     .L3

$ LANG=C gcc-3.4 -v
Reading specs from /usr/lib/gcc/i486-linux-gnu/3.4.6/specs
Configured with: ../src/configure -v --enable-languages=c,c++,f77,pascal
--prefix=/usr --libexecdir=/usr/lib --with-gxx-include-dir=/usr/include/c++/3.4
--enable-shared --with-system-zlib --enable-nls --without-included-gettext
--program-suffix=-3.4 --enable-__cxa_atexit --enable-clocale=gnu
--enable-libstdcxx-debug --with-tune=i686 i486-linux-gnu
Thread model: posix
gcc version 3.4.6 (Debian 3.4.6-1)

$ LANG=C gcc-4.1 -v
Using built-in specs.
Target: i486-linux-gnu
Configured with: ../src/configure -v
--enable-languages=c,c++,java,fortran,objc,obj-c++,ada,treelang --prefix=/usr
--enable-shared --with-system-zlib --libexecdir=/usr/lib
--without-included-gettext --enable-threads=posix --enable-nls
--program-suffix=-4.1 --enable-__cxa_atexit --enable-clocale=gnu
--enable-libstdcxx-debug --enable-java-awt=gtk --enable-gtk-cairo
--with-java-home=/usr/lib/jvm/java-1.4.2-gcj-4.1-1.4.2.0/jre --enable-mpfr
--with-tune=i686 --enable-checking=release i486-linux-gnu
Thread model: posix
gcc version 4.1.0 (Debian 4.1.0-1)


-- 
           Summary: Structures are copied byte by byte into function
                    arguments
           Product: gcc
           Version: 4.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: guillaume dot melquiond at ens-lyon dot fr
GCC target triplet: i486-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27055

[Bug c/27055] Structures are copied byte by byte into function arguments

[gdr at integrable-solutions dot net]

------- Comment #1 from gdr at integrable-solutions dot net  2006-04-06 12:03 -------
Subject: Re:   New: Structures are copied byte by byte into function arguments

"guillaume dot melquiond at ens-lyon dot fr" <gcc-bugzilla@gcc.gnu.org> writes:

[...]

| With 3.4.6, the copy is done byte-by-byte without string
| opcodes. With 4.0.3 and 4.1.0, it is done byte-by-byte and
| out-of-line: there are two jumps for each copied byte! So argument
| copy got broken for x86 during GCC 3.4 cycle and it did not get any
| better with GCC 4. 

That is insane, if I may.

-- Gaby


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27055

[Bug c/27055] Structures are copied byte by byte into function arguments

[pluto at agmk dot net]

------- Comment #2 from pluto at agmk dot net  2006-04-06 12:49 -------
for -march=i{345}86 I get:

f:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $4008, %esp
        movl    %esp, %eax
        pushl   %edx
        pushl   $4000
        pushl   8(%ebp)
        pushl   %eax
        call    memcpy
        addl    $16, %esp
        call    g
        addl    $4000, %esp
        leave
        ret

for -march=i686,pentium2,... I get:

f:
        pushl   %ebp
        xorl    %edx, %edx
        movl    %esp, %ebp
        subl    $4008, %esp
        movl    8(%ebp), %ecx
.L3:
        cmpl    $4000, %edx
        jb      .L2
        call    g
        leave
        ret
        .p2align 4,,7
.L2:
        movzbl  (%ecx,%edx), %eax
        movb    %al, (%esp,%edx)
        incl    %edx
        .p2align 4,,3
        jmp     .L3


$ gcc version 4.1.1 20060405 (prerelease) (PLD-Linux)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27055

[Bug c/27055] Structures are copied byte by byte into function arguments

[guillaume dot melquiond at ens-lyon dot fr]

------- Comment #3 from guillaume dot melquiond at ens-lyon dot fr  2006-04-06 13:05 -------
> for -march=i{345}86 I get
> for -march=i686,pentium2,... I get

Thanks to your tests, I noticed that the change of behavior between GCC 3.4
versions was actually caused by the addition of --with-tune=686 in Debian
packages. In fact, even older versions of GCC 3.4 are doing byte-by-byte copies
once you change the -march target to at least i686. GCC 3.3 is fine though.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27055

[Bug target/27055] Structures are copied byte by byte into function arguments

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2006-04-13 01:52 -------
This works on most targets I tried, ppc-linux-gnu and x86_64-linux-gnu both
work, only i?86-linux-gnu (32bit) did not.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
          Component|middle-end                  |target
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2006-04-13 01:52:19
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27055

[Bug target/27055] Structures are copied byte by byte into function arguments

[guillaume dot melquiond at ens-lyon dot fr]

------- Comment #5 from guillaume dot melquiond at ens-lyon dot fr  2006-04-16 18:15 -------
I have tried to look at why it fails. Here are my observations. The function
gcc/expr.c:emit_block_move tries to expand a movmem pattern. The backend
refuses, since a call to memcpy would be a lot more efficient. So
emit_block_move tries to generate memcpy but
block_move_libcall_safe_for_call_parm says this is not safe. The only thing
left is calling emit_block_move_via_loop, and it generates some really dumb
code.

Whenever there is -Os or -mtune=i386 or -mregparm=3, the function
block_move_libcall_safe_for_call_parm says memcpy is safe. Otherwise it fails
because arguments are neither pushed nor passed through registers but stored at
the top of the stack. Failing then seems to be the right thing to do, since the
call to memcpy will otherwise overwrite its own argument stack.

Assuming memcpy is the right thing to do, emit_block_move should allocate a bit
of stack space for memcpy instead of refusing to call it (and hence falling
back to a loop move).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27055