OpenE2K/gcc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

genconfig.c (main): Disable HAVE_conditional_arithmetic.

Browse Source 
* genconfig.c (main): Disable HAVE_conditional_arithmetic.
        * jump.c (jump_optimize_1): Document why.

From-SVN: r29841
This commit is contained in:
Richard Henderson 1999-10-06 10:22:54 -07:00 committed by Richard Henderson
parent f428f25240
commit 2ee2956a63
3 changed files with 39 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
gcc/ChangeLog
View File

@ -1,3 +1,8 @@
Wed Oct 6 10:21:15 1999 Richard Henderson <rth@cygnus.com>
* genconfig.c (main): Disable HAVE_conditional_arithmetic.
* jump.c (jump_optimize_1): Document why.
Wed Oct 6 10:41:56 1999 Kaveh R. Ghazi <ghazi@caip.rutgers.edu>
* collect2.c (scan_prog_file, scan_libraries): Prototype function

3
gcc/genconfig.c
View File

@ -348,8 +348,11 @@ from the machine description file `md'. */\n\n");
if (have_cmove_flag)
printf ("#define HAVE_conditional_move\n");
#if 0
/* Disabled. See the discussion in jump.c. */
if (have_cond_arith_flag)
printf ("#define HAVE_conditional_arithmetic\n");
#endif
if (have_lo_sum_flag)
printf ("#define HAVE_lo_sum\n");

31
gcc/jump.c
View File

@ -943,6 +943,37 @@ jump_optimize_1 (f, cross_jump, noop_moves, after_regscan, mark_labels_only)
#endif /* HAVE_cc0 */
#ifdef HAVE_conditional_arithmetic
/* ??? This is disabled in genconfig, as this simple-minded
transformation can incredibly lengthen register lifetimes.
Consider this example from cexp.c's yyparse:
234 (set (pc)
(if_then_else (ne (reg:DI 149) (const_int 0 [0x0]))
(label_ref 248) (pc)))
237 (set (reg/i:DI 0 $0) (const_int 1 [0x1]))
239 (set (pc) (label_ref 2382))
248 (code_label ("yybackup"))
This will be transformed to:
237 (set (reg/i:DI 0 $0)
(if_then_else:DI (eq (reg:DI 149) (const_int 0 [0x0]))
(const_int 1 [0x1]) (reg/i:DI 0 $0)))
239 (set (pc)
(if_then_else (eq (reg:DI 149) (const_int 0 [0x0]))
(label_ref 2382) (pc)))
which, from this narrow viewpoint looks fine. Except that
between this and 3 other ocurrences of the same pattern, $0
is now live for basically the entire function, and we'll
get an abort in caller_save.
Any replacement for this code should recall that a set of
a register that is not live need not, and indeed should not,
be conditionalized. Either that, or delay the transformation
until after register allocation. */
/* See if this is a conditional jump around a small number of
instructions that we can conditionalize. Don't do this before
the initial CSE pass or after reload.