[Bug middle-end/27016] ARM optimizer produces severely suboptimal code

["steven at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #3 from steven at gcc dot gnu dot org  2009-06-30 11:35 -------
For this test case:

unsigned int code_in_ram[100];

void testme(void)
{
  unsigned int *p_rom, *p_ram, *p_end, len;

  extern unsigned int _ram_erase_sector_start;
  extern unsigned int _ram_erase_sector_end;

  p_ram = code_in_ram;
  p_rom = &_ram_erase_sector_start;
  len = ((unsigned int)&_ram_erase_sector_end
         - (unsigned int)&_ram_erase_sector_start) / sizeof(unsigned int);

  for (len = 100; len > 0; --len)
    {
      *p_ram++ = *p_rom++;
    }
}


I get the following code with a 4.5.0 checkout from a few days ago (see
.ident):

        .cpu arm7tdmi
        .fpu softvfp
        .eabi_attribute 20, 1
        .eabi_attribute 21, 1
        .eabi_attribute 23, 3
        .eabi_attribute 24, 1
        .eabi_attribute 25, 1
        .eabi_attribute 26, 1
        .eabi_attribute 30, 4
        .eabi_attribute 18, 4
        .file   "t.c"
        .text
        .align  2
        .global testme
        .type   testme, %function
testme:
        @ Function supports interworking.
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        ldr     r1, .L5
        ldr     r2, .L5+4
        mov     r3, #0
.L2:
        ldr     r0, [r2, r3]
        str     r0, [r1, r3]
        add     r3, r3, #4
        cmp     r3, #400
        bne     .L2
        bx      lr
.L6:
        .align  2
.L5:
        .word   code_in_ram
        .word   _ram_erase_sector_start
        .size   testme, .-testme
        .comm   code_in_ram,400,4
        .ident  "GCC: (GNU) 4.5.0 20090626 (experimental) [trunk revision
148960]"


This looks marginally better than the code of comment #2.



With the original test case of comment #0 I get the following code:

        .cpu arm7tdmi
        .fpu softvfp
        .eabi_attribute 20, 1
        .eabi_attribute 21, 1
        .eabi_attribute 23, 3
        .eabi_attribute 24, 1
        .eabi_attribute 25, 1
        .eabi_attribute 26, 1
        .eabi_attribute 30, 4
        .eabi_attribute 18, 4
        .file   "t.c"
        .text
        .align  2
        .global testme
        .type   testme, %function
testme:
        @ Function supports interworking.
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        ldr     r2, .L5
        ldr     r3, .L5+4
        ldr     r1, .L5+8
        b       .L2
.L3:
        ldr     r0, [r3], #4
        str     r0, [r2, #-4]
.L2:
        cmp     r3, r1
        add     r2, r2, #4
        bcc     .L3
        bx      lr
.L6:
        .align  2
.L5:
        .word   code_in_ram
        .word   _ram_erase_sector_start
        .word   _ram_erase_sector_end
        .size   testme, .-testme
        .comm   code_in_ram,400,4
        .ident  "GCC: (GNU) 4.5.0 20090626 (experimental) [trunk revision
148960]"


The "ldr     r3, .L6+8" line of comment #0 is now hoisted from the loop into r1
in "ldr     r1, .L5+8".  GCC still does not use a post-increment for the store
in the loop (see use of r2), even though it uses a post-increment for the load.

So, bug still is there.  Might be a dup of one of those many other bugs about
poor use of auto-increment instructions by GCC... ;-)


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2009-06-30 11:35:26
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27016