[thumb2, size] Replace load/store by memcpy more aggressively

[Yao Qi]

struct record
{
   int Int_Comp;
   /* Increase array size by 1, load/store instructions will be replaced by memcpy(). */
   char Str_Comp [60];
};
extern void foo(strcut record* t);
void Proc_1 (struct record* Ptr_Val_Par)
{
  struct record rt = *Ptr_Val_Par;
  foo (&rt);
}

$ arm-none-linux-gnueabi-gcc -mthumb
 -c -mthumb -mcpu=cortex-a9 -mfpu=neon -Os -fno-common -mfloat-abi=hard -c -o 1.o 1.c

Generate code like this,
00000000 <Proc_1>:
   0: b530 push {r4, r5, lr}
   2: 4605 mov r5, r0
   4: b091 sub sp, #68 ; 0x44
   6: 466c mov r4, sp
   8: cd0f ldmia r5!, {r0, r1, r2, r3}
   a: c40f stmia r4!, {r0, r1, r2, r3}
   c: cd0f ldmia r5!, {r0, r1, r2, r3}
   e: c40f stmia r4!, {r0, r1, r2, r3}
  10: cd0f ldmia r5!, {r0, r1, r2, r3}
  12: c40f stmia r4!, {r0, r1, r2, r3}
  14: e895 000f ldmia.w r5, {r0, r1, r2, r3}
  18: e884 000f stmia.w r4, {r0, r1, r2, r3}
  1c: 4668 mov r0, sp
  1e: f7ff fffe bl 0 <foo>
  22: b011 add sp, #68 ; 0x44
  24: bd30 pop {r4, r5, pc}

However, when I increase the size of struct record, gcc will generate memcpy() rather than ldmia/stmia. Problem here is that we should make this more aggressive, in order to improve code size.

[Chung-Lin Tang (cltang)]

This form of block move is generated by arm_gen_movmemqi(), called from the "movmemqi" expand pattern.
See the beginning of that function in arm.c, at the length value in operands[2], and you should know what to do.

Adjusting the threshold value from 64 to 16, I got:
        push {lr}
        sub sp, sp, #68
        mov r1, r0
        movs r2, #64
        mov r0, sp
        bl memcpy
        mov r0, sp
        bl foo
        add sp, sp, #68
        pop {pc}

[Yao Qi (yao-codesourcery)]

Chung-Lin,
Thanks for this quick fix here. From the size optimization point of view, fix could be "reduce threshold value once -Os is turned on". Since we are in investigation step, I'll leave it here, and come back to fix it when we move to "implementation" step.