chp
Joined: 23 Jun 2004
Posts: 313
|
 Posted: Tue Feb 07, 2006 9:22 am Post subject: EE-GCC and __restrict |
 |
|
This patch will enable __restrict to work on functions that are inlined. A simple example on code that will build into more proper code is the following:
| Code: |
static inline sillyCopy(int* __restrict dest, const int* __restrict src, int size)
{
for (int i = 0; i < size; ++i)
{
dest[i+0] = src[i+0];
dest[i+1] = src[i+1];
dest[i+2] = src[i+0];
dest[i+3] = src[i+1];
}
}
int* data;
int* data2;
void main()
{
sillyCopy(data2,data,200);
}
(compile with: ee-gcc -O2 -S -std=c99 -o out.s in.c)
|
Without the patch, it will consider the parameters to be aliased and read data for every operation. With the patch, it will properly consider the parameters as not being aliased against each other and only fetch data twice for every four writes.
Oh, and the patch in question. This patch is derived from looking on why the cell-toolchain managed to inline code that normal GCC didn't.
| Code: |
--- gcc-3.2.2-base/gcc/integrate.c 2003-02-03 21:56:29.000000000 +0100
+++ gcc-3.2.2/gcc/integrate.c 2006-02-06 18:10:07.000000000 +0100
@@ -358,12 +358,18 @@
/* Copy the declaration. */
if (TREE_CODE (decl) == PARM_DECL || TREE_CODE (decl) == RESULT_DECL)
{
+ tree type;
+ type = TREE_TYPE(decl);
+
/* For a parameter, we must make an equivalent VAR_DECL, not a
new PARM_DECL. */
copy = build_decl (VAR_DECL, DECL_NAME (decl), TREE_TYPE (decl));
TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
TREE_READONLY (copy) = TREE_READONLY (decl);
TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
+
+ if (TYPE_RESTRICT(type))
+ DECL_POINTER_ALIAS_SET (copy) = -2;
}
else
{
|
This simple fix doesn't work on the PSP-toolchain however, since restricted pointers seems to be broken 4.0.2.
_________________
GE Dominator |
|
FAQ
Search
Memberlist
Usergroups
Register
Profile
Log in to check your private messages
Log in
