This is the mail archive of the gcc-bugs@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[Bug rtl-optimization/35281] [4.3/4.4 Regression] multiply with 0 generated for 64*32->64



------- Comment #7 from ubizjak at gmail dot com  2008-03-05 11:10 -------
It looks that:

2006-11-03  Paolo Bonzini  <bonzini@gnu.org>
            Steven Bosscher  <steven@gcc.gnu.org>

        * fwprop.c: New file.
        ...
        * cse.c (fold_rtx_subreg, fold_rtx_mem, fold_rtx_mem_1, find_best_addr,
        canon_for_address, table_size): Remove.
        (new_basic_block, insert, remove_from_table): Remove references to
        table_size.
        (fold_rtx): Process SUBREGs and MEMs with equiv_constant, make
        simplification loop more straightforward by not calling fold_rtx
        recursively.
        (equiv_constant): Move here a small part of fold_rtx_subreg,
        do not call fold_rtx.  Call avoid_constant_pool_reference
        to process MEMs.

removed this functionality. Specifically this part:

-    /* If this is a constant pool reference, we can fold it into its
-       constant to allow better value tracking.  */
-    if (base && GET_CODE (base) == SYMBOL_REF
-       && CONSTANT_POOL_ADDRESS_P (base))
-      {
-       rtx constant = get_pool_constant (base);
-       enum machine_mode const_mode = get_pool_mode (base);
-       rtx new;
-
-       if (CONSTANT_P (constant) && GET_CODE (constant) != CONST_INT)
-         {
-           constant_pool_entries_cost = COST (constant);
-           constant_pool_entries_regcost = approx_reg_cost (constant);
-         }
-
-       /* If we are loading the full constant, we have an
-          equivalence.  */
-       if (offset == 0 && mode == const_mode)
-         return constant;
-
-       /* If this actually isn't a constant (weird!), we can't do
-          anything.  Otherwise, handle the two most common cases:
-          extracting a word from a multi-word constant, and
-          extracting the low-order bits.  Other cases don't seem
-          common enough to worry about.  */
-       if (! CONSTANT_P (constant))
-         return x;
-
-       if (GET_MODE_CLASS (mode) == MODE_INT
-           && GET_MODE_SIZE (mode) == UNITS_PER_WORD
-           && offset % UNITS_PER_WORD == 0
-           && (new = operand_subword (constant,
-                                      offset / UNITS_PER_WORD,
-                                      0, const_mode)) != 0)
-         return new;
-
-       if (((BYTES_BIG_ENDIAN
-             && offset == GET_MODE_SIZE (GET_MODE (constant)) - 1)
-            || (! BYTES_BIG_ENDIAN && offset == 0))
-           && (new = gen_lowpart (mode, constant)) != 0)
-         return new;
-      }

This also explains why 4.2 doesn't fold HImode references ("Other cases don't
seem common enough to worry about.").

Can we have this functionality back, perhaps also for "Other cases", since
there are people that worry about them?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35281


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]