------- Comment #8 from ubizjak at gmail dot com 2008-04-25 09:55 -------
The problem is indeed in trisolve:
subroutine trisolve(x,i1,i2)
integer :: i1 , i2
real(dpkind),dimension(i2)::x
integer :: i
x(i1) = gi(i1)* x(i1)
do i = i1+1 , i2
x(i) = gi(i)*(x(i)-au1(i-1)*x(i-1))
enddo
do i = i2-1 , i1 , -1
x(i) = x(i) - gi(i)*au1(i)*x(i+1)
enddo
end subroutine trisolve
Please note two very tight loops that calculate x[n] from the value x[n-1],
where x[n-1] is the result of a previous step.
.127t.optimized tree dump for the the first loop (the second loop is the same,
only going from last to first element) in non-regressed case shows:
<bb 4>:
MEM[base: ivtmp.297] = MEM[base: ivtmp.295] * ((MEM[base: ivtmp.297] -
MEM[base: ivtmp.300] * MEM[base: ivtmp.297, offset: 0x0fffffffffffffff8]));
ivtmp.295 = ivtmp.295 + D.3347;
ivtmp.297 = ivtmp.297 + 8;
ivtmp.300 = ivtmp.300 + 8;
ivtmp.304 = ivtmp.304 + 1;
if ((integer(kind=4)) ivtmp.304 == D.1652)
goto <bb 5>;
else
goto <bb 4>;
this code results in:
.L3:
movsd (%r9), %xmm10
addl $4, %edx
movsd (%rcx), %xmm9
X+> mulsd -8(%rcx), %xmm10
movsd 8(%rcx), %xmm7
movsd 16(%rcx), %xmm5
movsd 24(%rcx), %xmm3
subsd %xmm10, %xmm9
mulsd (%rax), %xmm9
addq %r10, %rax
1-> movsd %xmm9, (%rcx)
movsd 8(%r9), %xmm8
1+> mulsd %xmm9, %xmm8
subsd %xmm8, %xmm7
mulsd (%rax), %xmm7
addq %r10, %rax
2-> movsd %xmm7, 8(%rcx)
movsd 16(%r9), %xmm6
2+> mulsd %xmm7, %xmm6
subsd %xmm6, %xmm5
mulsd (%rax), %xmm5
addq %r10, %rax
3-> movsd %xmm5, 16(%rcx)
movsd 24(%r9), %xmm4
addq $32, %r9
3+> mulsd %xmm5, %xmm4
subsd %xmm4, %xmm3
mulsd (%rax), %xmm3
addq %r10, %rax
X-> movsd %xmm3, 24(%rcx)
addq $32, %rcx
cmpl %ebp, %edx
jne .L3
In the code above, it can be seen how unrolled iterations are linked together.
The result from previous iteration (marked with N->) enters next iteration
(marked with N+>).
BTW: Optimizer could also link X-> and X+> but this is probably too much...
Patched gcc regressed in this area:
<bb 4>:
MEM[base: ivtmp.297] = MEM[base: ivtmp.295] * ((MEM[base: ivtmp.297] -
MEM[base: ivtmp.300] * MEM[base: ivtmp.302]));
ivtmp.295 = ivtmp.295 + D.3349;
ivtmp.297 = ivtmp.297 + 8;
ivtmp.300 = ivtmp.300 + 8;
ivtmp.302 = ivtmp.302 + 8;
ivtmp.304 = ivtmp.304 + 1;
if ((integer(kind=4)) ivtmp.304 == D.1652)
goto <bb 5>;
else
goto <bb 4>;
this code results in:
.L3:
movsd (%r9), %xmm10
addl $4, %edx
movsd (%rcx), %xmm9
X-> mulsd (%r8), %xmm10
movsd 8(%rcx), %xmm7
movsd 16(%rcx), %xmm5
movsd 24(%rcx), %xmm3
subsd %xmm10, %xmm9
mulsd (%rax), %xmm9
addq %rbx, %rax
1-> movsd %xmm9, (%rcx)
movsd 8(%r9), %xmm8
1+> mulsd 8(%r8), %xmm8
subsd %xmm8, %xmm7
mulsd (%rax), %xmm7
addq %rbx, %rax
2-> movsd %xmm7, 8(%rcx)
movsd 16(%r9), %xmm6
2+> mulsd 16(%r8), %xmm6
subsd %xmm6, %xmm5
mulsd (%rax), %xmm5
addq %rbx, %rax
3-> movsd %xmm5, 16(%rcx)
movsd 24(%r9), %xmm4
addq $32, %r9
3+> mulsd 24(%r8), %xmm4
addq $32, %r8
subsd %xmm4, %xmm3
mulsd (%rax), %xmm3
addq %rbx, %rax
X-> movsd %xmm3, 24(%rcx)
addq $32, %rcx
cmpl %r12d, %edx
jne .L3
In the code above, the links are broken. In ".+>" case, gcc reloads from memory
the same value that is otherwise available in the register, marked with ".->".
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34163
