http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49121
--- Comment #3 from vries at gcc dot gnu.org 2011-05-23 16:07:04 UTC ---
test case:
...
#ifndef TYPE
#define TYPE char*
#endif
extern int a[];
/* Can not infer loop iteration from array -- exit test can not be replaced.
*/
void foo (int i_width, TYPE dst, TYPE src1, TYPE src2)
{
TYPE dstn= dst + i_width;
TYPE dst0 = dst;
unsigned long long i = 0;
for( ; dst <= dstn; )
{
dst0[i] = ( src1[i] + src2[i] + 1 +a[i]) >> 1;
dst++;
i += 16;
}
}
...
test case representation at ivopts:
...
foo (int i_width, char * dst, char * src1, char * src2)
{
long long unsigned int i;
char * dstn;
char D.2707;
int D.2706;
int D.2705;
int D.2704;
int D.2703;
int D.2702;
int D.2701;
char D.2700;
char * D.2699;
int D.2698;
char D.2697;
char * D.2696;
char * D.2695;
long unsigned int D.2694;
<bb 2>:
D.2694_4 = (long unsigned int) i_width_3(D);
dstn_6 = dst_5(D) + D.2694_4;
if (dst_5(D) <= dstn_6)
goto <bb 3>;
else
goto <bb 7>;
<bb 3>:
<bb 4>:
# dst_32 = PHI <dst_24(5), dst_5(D)(3)>
# i_33 = PHI <i_25(5), 0(3)>
D.2695_9 = dst_5(D) + i_33;
D.2696_11 = src1_10(D) + i_33;
D.2697_12 = *D.2696_11;
D.2698_13 = (int) D.2697_12;
D.2699_15 = src2_14(D) + i_33;
D.2700_16 = *D.2699_15;
D.2701_17 = (int) D.2700_16;
D.2702_18 = D.2698_13 + D.2701_17;
D.2703_19 = D.2702_18 + 1;
D.2704_20 = a[i_33];
D.2705_21 = D.2703_19 + D.2704_20;
D.2706_22 = D.2705_21 >> 1;
D.2707_23 = (char) D.2706_22;
*D.2695_9 = D.2707_23;
dst_24 = dst_32 + 1;
i_25 = i_33 + 16;
if (dstn_6 >= dst_24)
goto <bb 5>;
else
goto <bb 6>;
<bb 5>:
goto <bb 4>;
<bb 6>:
<bb 7>:
return;
}
...
the patch causes estimated_loop_iterations (loop, true, &max_niter) in
may_eliminate_iv to return a different value:
without patch: 0x10000000000000000
with patch: 0x01000000000000000
The analysis in the patch concludes that the pointer arithmetic statement
D.2695_9 = dst_5(D) + i_33
can only be executed 0x01000000000000000 times, based on the fact that i is
incremented with 16 each iteration. Note that the statement corresponds to the
pointer arithmetic in 'dst0[i]'.
Since the proposed new char pointer iterator can represent 0x10000000000000000
distinct values, may_eliminate_iv returns true, and the exit test is replaced:
...
Replacing exit test: if (dstn_6 >= dst_24)
...
This causes the check:
...
/* { dg-final { scan-tree-dump-times "Replacing" 0 "ivopts"} } */
...
to fail.
The test case needs to be adapted.
