Re: [patch] Fix PR44483: incrementally gimplify BB predicates to avoid redundant computations.

[Ira Rosen <IRAR at il dot ibm dot com>]

Hi,

gcc-patches-owner@gcc.gnu.org wrote on 11/06/2010 01:52:58 AM:
>
> Hi,
>
> The attached patch fixes PR44483 by early gimplifying the expressions
> that are used in the predicates of basic blocks: in the testcase,
> these expressions are redundant and if not gimplified in time, they
> end up to form an exponential tree of conditions.
>
> With this patch we gimplify the predicates, and collect the gimple
> sequences during the analysis phase.  Because we now force the
> predicate of a BB to be an SSA_NAME, this is what is used further on
> for the next basic block, avoiding the full expression to be
> recomputed.  When the if-conversion analysis fails, the SSA_NAMEs used
> in the gimple sequences are released using free_stmt_operands.
> Otherwise the gimple sequences are inserted in the end of each basic
> block that is predicated by these computations.  These computations
> are then used in the if-conversion of the PHI nodes following the BB.


With this patch if-conversion generates for

  for (i = 0; i < N; i++)
    if (arr[i] < limit)
      {
        pos = i + 1;
        limit = arr[i];
      }

some redundant statements, that were not created before the patch (none of
them are used outside the loop):

  # pos_22 = PHI <pos_1(4), 1(2)>
  # i_23 = PHI <prephitmp.8_2(4), 0(2)>
  # limit_24 = PHI <limit_4(4), 1.28e+2(2)>
  # ivtmp.9_18 = PHI <ivtmp.9_17(4), 64(2)>
  limit_9 = arr[i_23];
  pos_10 = i_23 + 1;
  D.4534_12 = limit_9 < limit_24;                       <-----
  pretmp.7_3 = i_23 + 1;
  D.4535_20 = limit_9 >= limit_24;                      <-----
  pos_1 = [cond_expr] limit_9 >= limit_24 ? pos_22 : pos_10;
  limit_4 = [cond_expr] limit_9 >= limit_24 ? limit_24 : limit_9;
  prephitmp.8_2 = [cond_expr] limit_9 >= limit_24 ? pretmp.7_3 : pos_10;
  ivtmp.9_17 = ivtmp.9_18 - 1;
  D.4536_19 = D.4534_12 || D.4535_20;                   <-----
  if (ivtmp.9_17 != 0)
    goto <bb 4>;
  else
    goto <bb 5>;

The statements are removed by later dce pass, but they interfere with my
attempts to vectorize this loop.
Is it possible to get rid of them in if-conversion? Should I open a PR?

Thanks,
Ira


>
> The patch passed make -k check RUNTESTFLAGS=tree-ssa.exp and vect.exp
> and is right now under regstrap on amd64-linux.  Ok for trunk after
regstrap?
>
> Thanks,
> Sebastian Pop
> --
> AMD / Open Source Compiler Engineering / GNU Tools
> [attachment "0001-Fix-PR44483-incrementally-gimplify-BB-predicates-
> to-.patch" deleted by Ira Rosen/Haifa/IBM]