https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68576
--- Comment #3 from majun <majun4950646 at 163 dot com> ---
(In reply to Richard Biener from comment #2)
> I think the issue is that with respect to loop 4 the evolution is
>
> _76 = { 1 + _373, + , 1 }_4
>
> but with all the casting we end up with
>
> (set_scalar_evolution
> instantiated_below = 5
> (scalar = _470)
> (scalar_evolution = {(unsigned long) stride.92_29 + (unsigned long)
> offset.93_35, +, (unsigned long) stride.92_29}_1))
> )
>
> and instantiate_scev for the SCEV above fails. Likewise for instantiate
> below 10, only below 15 succeeds in the end.
Yes,it failed in CASE_CONVERT while do instantiate_scev_r.
BTW, another example is interchange-1.f:
---
subroutine foo(f1,f2,f3,f4,f5,f6,f7,f8,f9,f0,g1,g2,g3)
implicit none
integer f4,f3,f2,f1
integer g4,g5,g6,g7,g8,g9
integer i1,i2,i3,i4,i5
real*8 g1(5,f3,f2,f1),g2(5,5,f3,f2,f1),g3(5,f3,f2,f1)
real*8 f0(5,5,f3,f2,f1),f9(5,5,f3,f2,f1),f8(5,5,f3,f2,f1)
real*8 f7(5,5,f3,f2,f1),f6(5,5,f3,f2,f1),f5(5,5,f3,f2,f1)
do i3=1,f1
g8=mod(i3+f1-2,f1)+1
g9=mod(i3,f1)+1
do i4=1,f2
g6=mod(i4+f2-2,f2)+1
g7=mod(i4,f2)+1
do i5=1,f3
g4=mod(i5+f3-2,f3)+1
g5=mod(i5,f3)+1
do i1=1,5
g3(i1,i5,i4,i3)=0.0d0
do i2=1,5
g3(i1,i5,i4,i3)=g3(i1,i5,i4,i3)+
1 g2(i1,i2,i5,i4,i3)*g1(i2,i5,i4,i3)+
2 f0(i1,i2,i5,i4,i3)*g1(i2,g5,i4,i3)+
3 f9(i1,i2,i5,i4,i3)*g1(i2,i5,g7,i3)+
4 f8(i1,i2,i5,i4,i3)*g1(i2,i5,i4,g9)+
5 f7(i1,i2,i5,i4,i3)*g1(i2,g4,i4,i3)+
6 f6(i1,i2,i5,i4,i3)*g1(i2,i5,g6,i3)+
7 f5(i1,i2,i5,i4,i3)*g1(i2,i5,i4,g8)
enddo
enddo
enddo
enddo
enddo
return
end
---
the innermost 2 loops have constant loop bound, and this example can paralle in
the 2nd loop instead of 3rd loop.
so,any plan to improve CASE_CONVERT or type casting in instantiate_scev ?
Thanks!
---Jun
