https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65709
Paolo Bonzini <bonzini at gnu dot org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |bonzini at gnu dot org
--- Comment #20 from Paolo Bonzini <bonzini at gnu dot org> ---
> how to efficiently access unaligned memory ?
Use memcpy between unsigned char pointers and with a constant size. The
compiler knows to translate it to an unaligned memory access, or even a
combination of unaligned and aligned memory accesses:
$ cat f.c
int f(char *restrict a, const char *restrict b)
{
int i;
for (i = 0; i < 512; i++)
a[i] = b[i];
}
$ gcc f.c -O3 -S -o f.s -fdump-tree-optimized
$ cat f.c.191t.optimized
;; Function f (f, funcdef_no=0, decl_uid=1832, cgraph_uid=0, symbol_order=0)
f (char * restrict a, const char * restrict b)
{
<bb 2>:
__builtin_memcpy (a_5(D), b_7(D), 512); [tail call]
return;
}
$ cat f.s
...
f:
.cfi_startproc
movq (%rsi), %rdx
movq %rdi, %rax
leaq 8(%rdi), %rdi
movq %rdx, -8(%rdi)
movq 504(%rsi), %rdx
movq %rdx, 496(%rdi)
andq $-8, %rdi
subq %rdi, %rax
subq %rax, %rsi
addl $512, %eax
shrl $3, %eax
movl %eax, %ecx
rep movsq
ret
...
It's doing unaligned accesses for the first and last 8 bytes, and 31 aligned
8-byte accesses in the middle.
