[Bug tree-optimization/34683] [4.3 Regression] Fortran FE generated IL pessimizes middle-end IL and analysis

["rguenth at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #31 from rguenth at gcc dot gnu dot org  2008-01-08 12:55 -------
Another hot assert is

static inline tree
memory_partition (tree sym)
{
  tree tag;

  /* MPTs belong to their own partition.  */
  if (TREE_CODE (sym) == MEMORY_PARTITION_TAG)
    return sym;

  gcc_assert (!is_gimple_reg (sym));


and is_gimple_reg isn't cheap either.  For this testcase two thirds of
the calls are attributed to add_vars_for_offset (through add_v{use,def}):

                0.50    0.01 135234172/199610365     add_vars_for_offset [27]
[114]    1.1    0.74    0.01 199610365         is_gimple_reg [114]

and it really isn't cheap:

CPU: AMD64 processors, speed 2000 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit mask
of 0x00 (No unit mask) count 100000
samples  %        symbol name
545326   36.8318  ggc_alloc_stat
310838   20.9943  add_vars_for_offset
85655     5.7852  add_virtual_operand
69156     4.6709  finalize_ssa_stmt_operands
57057     3.8537  bitmap_set_bit
36755     2.4825  htab_find_with_hash
35915     2.4257  get_expr_operands
28207     1.9051  is_gimple_reg
23340     1.5764  referenced_var_lookup

wrapping the assert inside ENABLE_CHECKING buys us 2s or 2.5%.  But this
is too much micro-optimizing for me ;)

I'd rather have the quadratic SSA rewriting in the unroller fixed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34683