[Bug target/52338] New: shorter abs thumb2 code sequences

["carrot at google dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52338

             Bug #: 52338
           Summary: shorter abs thumb2 code sequences
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: carrot@google.com
            Target: arm-linux-androideabi


Compile the following code with options -march=armv7-a -mthumb -Os


char placex[] = {1, 2, 3, 4};
char placey[] = {5, 6, 7, 8};

int t0k(int b, int w)
{
   int h1 = placex[b] - placex[w];
   int h2 = placey[b] - placey[w];
   if (h1 < 0) h1 = -h1;
   if (h2 < 0) h2 = -h2;
   return h1 + h2;
}


GCC 4.7 generates:


t0k:
    ldr    r3, .L2
    push    {r4, lr}
    ldrb    r4, [r3, r0]    @ zero_extendqisi2
    adds    r0, r3, r0
    ldrb    r2, [r3, r1]    @ zero_extendqisi2
    adds    r3, r3, r1
    ldrb    r0, [r0, #4]    @ zero_extendqisi2
    ldrb    r3, [r3, #4]    @ zero_extendqisi2
    subs    r2, r4, r2
    subs    r3, r0, r3
    eor    r0, r2, r2, asr #31    //  abs(h1)
    sub    r0, r0, r2, asr #31    //  abs(h1)
    cmp    r3, #0                 //  abs(h2)
    it    lt                     //  abs(h2)
    rsblt    r3, r3, #0             //  abs(h2)
    adds    r0, r0, r3
    pop    {r4, pc}


It's interesting that gcc generats two different code sequence for abs(h1) and
abs(h2).

After carefully studying the insn pattern "*thumb2_abssi2", I got the
impression that the selection of code sequence is depends on if the result
register is same as source register. If they are same, the cmp code sequence is
used, otherwise the pure arithmetic code sequence is used.

But the cmp code sequence can also be used when the source is not the same as
target register, and may result in smaller code size, like in this case.

Another question is there any performance differences between these two
different code sequences?