Problem in SH code generation-mainline/BIB (probably other targets too)

["Naveen Sharma, Noida" <naveens at noida dot hcltech dot com>]

Hi,

There seems to be a problem in GCC code generation for SH4 and 
probably other targets.

As an example consider the  test program 
gcc.c-torture/execute/931004-13.c in gcc testsuite.

I am seeing the the following code.
Options : "-O2 -ml -m4 -fno-schedule-insns2"
gcc version 3.4-bi 20021104 (experimental)  

main ()
{
  .... Snip .....
  struct tiny x[3];
  x[0].c = 10;
  c8:   0a e1           mov     #10,r1
  ca:   10 2e           mov.b   r1,@r14
  x[1].c = 11;
  cc:   0b e2           mov     #11,r2
  ce:   e3 61           mov     r14,r1
  d0:   04 71           add     #4,r1
  d2:   20 21           mov.b   r2,@r1
  x[2].c = 12;
  d4:   0c e2           mov     #12,r2
  d6:   e3 61           mov     r14,r1   <--- This can be eliminated.
  d8:   08 71           add     #8,r1
  da:   20 21           mov.b   r2,@r1
     ............

Whereas with a previous snapshot (gcc version 3.2 20020503), 
I get 

main ()
{
................ 
  struct tiny x[3];
  x[0].c = 10;
  c8:   0a e1           mov     #10,r1
  ca:   10 2e           mov.b   r1,@r14
  x[1].c = 11;
  cc:   0b e2           mov     #11,r2
  ce:   e3 61           mov     r14,r1
  d0:   04 71           add     #4,r1
  d2:   20 21           mov.b   r2,@r1
  x[2].c = 12;
  d4:   0c e2           mov     #12,r2
  d6:   04 71           add     #4,r1     <---- Note this
  d8:   20 21           mov.b   r2,@r1
.......................

So, it seems that  the following transformation 
is not happening.

  pX <- pA + N
    ...
  pX <- pA + M
  into
  pX <- pA + N
  ...
  pX <- pX + (M - N)

This has led to significant increase in code size. As per my understanding 
either regmove pass or reload_cse_regs () would be doing this
transfromation.

I am investigating which patch is responsible. But if someone already knows 
it, it would be helpful.

Best Regards,
  Naveen Sharma.