The following two functions are equivalent (especially after loop unrolling):
void
foo(const int *restrict a, int *restrict b, int *restrict c)
{
b[0] += a[0];
c[0] += a[0];
b[1] += a[1];
c[1] += a[1];
}
void
bar(const int *restrict a, int *restrict b, int *restrict c)
{
for (int i = 0; i < 2; ++i)
{
b[i] += a[i];
c[i] += a[i];
}
}
However gcc forgets about 'restrict' after the first iteration of the loop, and
foo() and bar() produce different code:
foo:
pushl %ebx
movl 8(%esp), %ebx
movl 12(%esp), %eax
movl 16(%esp), %edx
movl (%ebx), %ecx
addl %ecx, (%eax)
addl %ecx, (%edx) ;; Correct: no reloading of (%ebx) is needed.
movl 4(%ebx), %ecx
addl %ecx, 4(%eax)
addl %ecx, 4(%edx) ;; Correct: no reloading of 4(%ebx) is needed.
popl %ebx
ret
bar:
pushl %ebx
movl 8(%esp), %ebx
movl 12(%esp), %edx
movl 16(%esp), %ecx
movl (%ebx), %eax
addl %eax, (%edx)
addl %eax, (%ecx) ;; Correct: no reloading of (%ebx) is needed.
movl 4(%ebx), %eax
addl %eax, 4(%edx)
movl 4(%ebx), %eax ;; BUG: unnecessary reloading of 4(%ebx).
addl %eax, 4(%ecx)
popl %ebx
ret
For any number of iterations only the first iteration honors the 'restrict'
qualifier. This is wrong, because 'restrict' is a property of a pointer, not
data, so if p and q pointers reference different objects, then (p + OFF1) and
(q + OFF2) also expected to reference different objects. Correct assembler for
foo() supports that.
--
Summary: 'restrict' is forgotten after loop unrolling
Product: gcc
Version: 4.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: tomash dot brechko at gmail dot com
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32273