Unexpected assembly generated with -Os

Hi,

/* BEGIN weird.c */
#include <stdint.h>

static const unsigned N = 42;

void clear( uint32_t x[N] )
{
    for( unsigned i = 0; i < N; i++ )
        x[i] = 0;
}
/* END weird.c */

% arm-none-eabi-gcc -std=c11 -march=armv7-m -mthumb weird.c -S -Os -o - | tail -n 14
clear:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        subs r3, r0, #4
        adds r0, r0, #164
.L2:
        movs r2, #0
        str r2, [r3, #4]!
        cmp r3, r0
        bne .L2
        bx lr
        .size clear, .-clear
        .ident "GCC: (GNU Tools for ARM Embedded Processors) 4.9.3 20141119 (release) [ARM/embedded-4_9-branch revision 218278]"

Observed behaviour: "mov r2, #0" is inside the loop, which hurts performance unless I'm really missing something.

Expected behaviour: this instruction should be outside the loop, yielding better performance without increasing the code size. (This is what we get with -O1 and -O2.)

Am I missing something big or is it a bug in some optimization pass enabled by -Os?

As a side note, in this case -O3 actually generates the most compact code for this function, ie a call to memset.

Thanks,
Manuel.