Re: arm memcpy of aligned data

[Kyrill Tkachov <kyrylo dot tkachov at arm dot com>]

Hi Mike,

On 28/05/15 22:15, Mike Stump wrote:
> So, the arm memcpy code of aligned data isn’t as good as it can be.
>
> void *memcpy(void *dest, const void *src, unsigned int n);
>
> void foo(char *dst, int i) {
>    memcpy (dst, &i, sizeof (i));
> }
>
> generates horrible code, but, it we are willing to notice the src or the destination are aligned, we can do much better:
>
> $ ./cc1 -fschedule-fusion -fdump-tree-all-all -da -march=armv7ve -mcpu=cortex-m4 -fomit-frame-pointer -quiet -O2 /tmp/t.c -o t.s
> $ cat t.s
> [ … ]
> foo:
> 	@ args = 0, pretend = 0, frame = 4
> 	@ frame_needed = 0, uses_anonymous_args = 0
> 	@ link register save eliminated.
> 	sub	sp, sp, #4
> 	str	r1, [r0]	@ unaligned
> 	add	sp, sp, #4

I think there's something to do with cpu tuning here as well.
For the code you've given compiled with -O2 -mcpu=cortex-a53 I get:
        sub     sp, sp, #8
        mov     r2, r0
        add     r3, sp, #8
        str     r1, [r3, #-4]!
        ldr     r0, [r3]        @ unaligned
        str     r0, [r2]        @ unaligned
        add     sp, sp, #8
        @ sp needed
        bx      lr

whereas for -O2 -mcpu=cortex-a57 I get the much better:
        sub     sp, sp, #8
        str     r1, [r0]        @ unaligned
        add     sp, sp, #8
        @ sp needed
        bx      lr

Kyrill


> Index: gcc/config/arm/arm.c
> ===================================================================
> --- gcc/config/arm/arm.c	(revision 223842)
> +++ gcc/config/arm/arm.c	(working copy)
> @@ -14376,7 +14376,10 @@ arm_block_move_unaligned_straight (rtx d
>   				srcoffset + j * UNITS_PER_WORD - src_autoinc);
>   	  mem = adjust_automodify_address (srcbase, SImode, addr,
>   					   srcoffset + j * UNITS_PER_WORD);
> -	  emit_insn (gen_unaligned_loadsi (regs[j], mem));
> +	  if (src_aligned)
> +	    emit_move_insn (regs[j], mem);
> +	  else
> +	    emit_insn (gen_unaligned_loadsi (regs[j], mem));
>   	}
>         srcoffset += words * UNITS_PER_WORD;
>       }
> @@ -14395,7 +14398,10 @@ arm_block_move_unaligned_straight (rtx d
>   				dstoffset + j * UNITS_PER_WORD - dst_autoinc);
>   	  mem = adjust_automodify_address (dstbase, SImode, addr,
>   					   dstoffset + j * UNITS_PER_WORD);
> -	  emit_insn (gen_unaligned_storesi (mem, regs[j]));
> +	  if (dst_aligned)
> +	    emit_move_insn (mem, regs[j]);
> +	  else
> +	    emit_insn (gen_unaligned_storesi (mem, regs[j]));
>   	}
>         dstoffset += words * UNITS_PER_WORD;
>       }
>
>
> Ok?
>
> Can someone spin this through an arm test suite run for me, I was doing this by inspection and cross compile on a system with no arm bits.  Bonus points if you can check it in with the test case above marked up as appropriate.