------- Additional Comments From steven at gcc dot gnu dot org 2005-09-18
14:27 -------
Not being even closely an RTL guru (as you can see in this bug audit trail ;-)
I'd propose the following patch:
Index: cse.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/cse.c,v
retrieving revision 1.360
diff -u -3 -p -r1.360 cse.c
--- cse.c 6 Sep 2005 08:52:51 -0000 1.360
+++ cse.c 18 Sep 2005 14:19:02 -0000
@@ -2999,7 +2999,8 @@ find_best_addr (rtx insn, rtx *loc, enum
p = p->next_same_value, count++)
if (! p->flag
&& (REG_P (p->exp)
- || exp_equiv_p (p->exp, p->exp, 1, false)))
+ || (GET_CODE (p->exp) != EXPR_LIST
+ && exp_equiv_p (p->exp, p->exp, 1, false))))
{
rtx new = simplify_gen_binary (GET_CODE (*loc), Pmode,
p->exp, op1);
Rationale:
We start with this:
(insn 16 15 20 0 (set (reg:DI 62)
(reg:DI 0 ax)) 81 {*movdi_1_rex64} (nil)
(insn_list:REG_RETVAL 14 (expr_list:REG_EQUAL (expr_list:REG_DEP_TRUE (use
(mem:BLK (scratch) [0 A8]))
(expr_list:REG_DEP_TRUE (symbol_ref:DI ("strlen") [flags 0x41]
<function_decl 0x2a9590d500 strlen>)
(expr_list:REG_DEP_TRUE (symbol_ref:DI ("savecallsin")
[flags 0x2] <var_decl 0x2a95a3cdc0 savecallsin>)
(nil))))
(nil))))
(insn 20 16 21 0 (set (reg:CCZ 17 flags)
(compare:CCZ (mem/s:QI (plus:DI (plus:DI (reg:DI 62)
(reg/f:DI 60))
(const_int -1 [0xffffffffffffffff])) [0 savecallsin S1
A8])
(const_int 47 [0x2f]))) 10 {*cmpqi_1} (nil)
(expr_list:REG_EQUAL (compare:CCZ (mem/s:QI (plus:DI (reg:DI 62)
(const:DI (plus:DI (symbol_ref:DI ("savecallsin") [flags
0x2] <var_decl 0x2a95a3cdc0 savecallsin>)
(const_int -1 [0xffffffffffffffff])))) [0
savecallsin S1 A8])
(const_int 47 [0x2f]))
(nil)))
and then we try to replace (reg:DI 62) with the rediculous EXPR_LIST. That
doesn't make any sense at all, to begin with (and the fact that we find that
the new address with the EXPR_LIST is cheaper is equally stupid). exp_equiv_p
is right that the canonicalized p->exp would be equal (nothing is volatile,
etc.). I guess the only proper way to prevent forming this new address is by
disallowing propagation of EXPR_LISTs completely. It never makes sense to
have an address containing an EXPR_LIST anyway.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23943
