https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64705

--- Comment #2 from amker at gcc dot gnu.org ---
Loop dump before IVOPT is like below:

Loop 4, basic blocks 28/30;

  <bb 26>:
  count_54 = count_172 + 1;
  _55 = i_161 + i_161;
  prime_56 = _55 + 3;
  k_57 = prime_56 + i_161;
  if (size_26 >= k_57)
    goto <bb 27>;
  else
    goto <bb 31>;

  <bb 27>:

  <bb 28>:
  # k_167 = PHI <k_57(27), k_62(30)>
  # ci_168 = PHI <ci_169(27), ci_58(30)>
  ci_58 = ci_168 + 1;
  k.19_59 = (sizetype) k_167;
  _60 = flags_30 + k.19_59;
  *_60 = 0;
  k_62 = prime_56 + k_167;
  if (size_26 >= k_62)
    goto <bb 30>;
  else
    goto <bb 29>;

  <bb 29>:
  # ci_154 = PHI <ci_58(28)>
  goto <bb 31>;

  <bb 30>:
  goto <bb 28>;


The IV uses found by IVOPT is like below:

use 0
  address
  defined in statement 
  used in statement *_60 = 0;

  at position *_60
  type char *
  base flags_30 + (sizetype) k_57
  step (sizetype) prime_56
  base object (void *) flags_30
  related candidates 
use 1
  compare
  defined in statement 
  used in statement if (size_26 >= k_62)

  at position 
  type long int
  base (_55 + 3) + k_57
  step prime_56
  is a biv
  related candidates 
use 2
  generic (computed on exit edge)
  defined in statement ci_58 = ci_168 + 1;
  used in statement ci_154 = PHI <ci_58(28)>

  at position 
  type long int
  base ci_169 + 1
  step 1
  is a biv
  related candidates 

Root cause is IVOPT expands use 1 from {prime_56 + k_57, prime_56}_loop to
{(_55 + _3) + k_57, prime_56}_loop.  Thus information of "iv.step == prime_56
== (_55+_3)" is lost during costs computation and uses rewrting, resulting in
wrong candidate selected and bloated loop after IVOPT.

The related code is in function find_givs_in_stmt_scev, specifically,

  if (!simple_iv (loop, loop_containing_stmt (stmt), lhs, iv, true))
    return false;
  iv->base = expand_simple_operations (iv->base);   // <--- expansion

I will see how to fix the issue by skipping expansion in case like this.

Thanks,
bin

Reply via email to