It appears that GCSE considers "read only memory" as call clobbered, which
is not the case in CSE. I have took the test for read-only memory from CSE
and add it to GCSE where we compute the transparency. Here is a patch that
does so. This patch makes gcse eliminate redundant loads after stores for
the following example. The difference is seen when we compile with the
options:
"-O3 --param max-gcse-passes=3" with/without "-fgcse-las" (on a
powerpc-linux target). If this looks a reasonable change I will
regression-test and bootstrap the patch and ask for commit.
Example:
#define CALL_FPTR(fptr) (*fptr)
#define MY_FOO_CHECK() if (my_foo_var) my_foo_func()
int my_foo_var;
struct my_foo_struct {
int my_dummy_field;
int *(*ppaddr)(int);
};
struct my_foo_struct *my_foo_record;
int my_main_foo(int n)
{
while ((my_foo_record = CALL_FPTR(my_foo_record->ppaddr)(n))) {
MY_FOO_CHECK();
}
return 0;
}
The patch:
2005-05-09 Mostafa Hagog <[EMAIL PROTECTED]>
* gcse.c (compute_transp): use MEM_READONLY_P in case of MEM.
Index: gcse.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/gcse.c,v
retrieving revision 1.340
diff -c -p -r1.340 gcse.c
*** gcse.c 23 Apr 2005 21:27:44 -0000 1.340
--- gcse.c 5 May 2005 12:54:04 -0000
*************** compute_transp (rtx x, int indx, sbitmap
*** 2470,2479 ****
do any list walking for them. */
EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls, 0, bb_index, bi)
{
! if (set_p)
! SET_BIT (bmap[bb_index], indx);
! else
! RESET_BIT (bmap[bb_index], indx);
}
/* Now iterate over the blocks which have memory modifications
--- 2470,2482 ----
do any list walking for them. */
EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls, 0, bb_index, bi)
{
! if (! MEM_READONLY_P (x))
! {
! if (set_p)
! SET_BIT (bmap[bb_index], indx);
! else
! RESET_BIT (bmap[bb_index], indx);
! }
}