Hi,
This question is about loop peeling done by gcc-4.1.2
in autovectorization.
For the following program gcc is able to do vectorize
the code by peeling.
1 #include<stdio.h>
2 int main()
3 {
4 int a[100],b[100],c[100];
5 int i,j,n=80;
6 for(i=0;i<100;i++) a[i]=b[i]=c[i]=i;
7
8
9 for(i=0;i<n;i++)
10 {
11 a[i+1]=b[i+13]-10;
12 b[i+1]=a[i+14]*3;
13 }
14
15 printf("\n %d %d", a[44],b[33]);
16 }
But, if we change the 12 th line to
b[i+2]=a[i+14]*3;
gcc cant do loop peeling.Consequently gcc cant do
autovectorization.
The reason is when the two references misalignments
are equal it sets DR_MISALIGNMENT(dr) to zero.
Below is the code from "tree-vect-analyze.c" where it
does so.
My question is why gcc cant do autovectorization in
case if two datarefernces's misalignments are not
same.
What should I do to improve that case?
Thanks in advance.
833 vect_update_misalignment_for_peel (struct
data_reference *dr,
834 struct
data_reference *dr_peel, int npeel)
835 {
836 unsigned int i;
837 int drsize;
838 VEC(dr_p,heap) *same_align_drs;
839 struct data_reference *current_dr;
840
841 if (known_alignment_for_access_p (dr)
842 && DR_MISALIGNMENT (dr) ==
DR_MISALIGNMENT (dr_peel))
843 {
844 DR_MISALIGNMENT (dr) = 0;
845 return;
846 }
___________________________________________________________
Yahoo! Answers - Got a question? Someone out there knows the answer. Try it
now.
http://uk.answers.yahoo.com/
