[Bug middle-end/78411] [7 Regression] FAIL: gcc.target/i386/pr45685.c scan-assembler-times cmov 6

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78411

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
diff --git a/gcc/tree-ssa-pre.c b/gcc/tree-ssa-pre.c
index 4f17200..c50c598 100644
--- a/gcc/tree-ssa-pre.c
+++ b/gcc/tree-ssa-pre.c
@@ -3017,6 +3017,7 @@ insert_into_preds_of_block (basic_block block, unsigned
int exprnum,
     }

   /* Make the necessary insertions.  */
+  bool all_constant = true;
   FOR_EACH_EDGE (pred, ei, block->preds)
     {
       gimple_seq stmts = NULL;
@@ -3041,8 +3042,15 @@ insert_into_preds_of_block (basic_block block, unsigned
int exprnum,
       if (is_gimple_min_invariant (builtexpr))
        avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
       else
-       avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
+       {
+         avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
+         all_constant = false;
+       }
     }
+  /* If we're about to insert a PHI with all arguments constant refuse
+     to do that.  But return whether we inserted stuff.  */
+  if (all_constant)
+    return insertions;
   /* If we didn't want a phi node, and we made insertions, we still have
      inserted new stuff, and thus return true.  If we didn't want a phi node,
      and didn't make insertions, we haven't added anything new, so return


but that still leaves us with

  _25 = MEM[base: products_16(D), index: i_35, step: 8, offset: 0B];
  if (_25 > 0)
    goto <bb 7>;
  else
    goto <bb 6>;

  <bb 6>:
  _17 = -_25;

  <bb 7>:
  # iftmp.0_13 = PHI <1(5), -1(6)>
  # prephitmp_10 = PHI <_25(5), _17(6)>
  _9 = (long long unsigned int) prephitmp_10;

and thus phiopt not matching the ABS.  For some reason split-paths split
the loop latch confusing RTL if-conversion even more (removing the
possibility to emit a conditional store).

I don't believe what tree-if-conv.c did was reasonable here (together with PRE)
but we do not generate optimal code from the code in the testcase.

Still if-conversion is generally a RTL opt task.