Hi,
I got a few size regressions when moving from 4.5.3 to 4.6.1, all due to
the same issue.
I have code that is basically a double word memory move:
void simple1(uint32 *a, uint32 *b) { *a = *b; }
GCC 4.6.1 is from this gimple:
simple1 (uint32 * a, uint32 * b)
{
uint32 D.1927;
# BLOCK 2 freq:10000
# PRED: ENTRY [100.0%] (fallthru,exec)
D.1927_2 = *b_1(D);
*a_3(D) = D.1927_2;
return;
# SUCC: EXIT [100.0%]
}
calling gen_movhi twice with:
call1 :
operand[0]: (reg:HI 19 [ D.1927 ])
operand[1]: (mem:HI (reg/v/f:QI 21 [ b ]))
call2 :
operand[0]: (mem:HI (reg/v/f:QI 20 [ a ]))
operand[1]: (reg:HI 19 [ D.1927 ])
while GCC 4.5.3 for _exactly_ the same gimple (as far as it is shown in
the logs) it only calls gen_movhi with:
operand[0]: (mem:HI (reg/v/f:QI 20 [ a ]))
operand[1]: (mem:HI (reg/v/f:QI 21 [ b ]))
This seems to boil down to code that looks exactly the same between
4.5.3-4.6.1 in cfgexpand.c, expand_gimple_basic_block:
def_operand_p def_p;
def_p = SINGLE_SSA_DEF_OPERAND (stmt, SSA_OP_DEF);
if (def_p != NULL)
{
/* Ignore this stmt if it is in the list of
replaceable expressions. */
if (SA.values
&& bitmap_bit_p (SA.values,
SSA_NAME_VERSION (DEF_FROM_PTR (def_p))))
continue;
}
last = expand_gimple_stmt (stmt);
maybe_dump_rtl_for_gimple_stmt (stmt, last);
gcc4.5.3 hits continue the first time it gets there and gcc4.6.1 fails
the inner if and enters expand_gimple_stmt twice.
From the comment in ssaexpand.h for the definition of SA.values, it says:
/* For an SSA name version V bit V is set iff TER decided that
its definition should be forwarded. */
bitmap values;
What is TER? Any hints on why GCC 4.6.1 has now a different behaviour to
GCC 4.5.3?
Cheers,
--
Paulo Matos
