--- Begin Message ---
|
Subject: |
Re: [Various] HAVE_FAST_UNALIGNED_ACCESS |
|
Date: |
Fri, 31 Mar 2023 16:26:13 +0800 |
Thanks, Robert. You also missed this one intended for Mattias:
From: Po Lu <luangruo@yahoo.com>
To: Mattias Engdegård <mattiase@acm.org>
Cc: Robert Pluim <rpluim@gmail.com>, emacs-devel@gnu.org
Subject: Re: HAVE_FAST_UNALIGNED_ACCESS
In-Reply-To: <87A3D742-5B8B-4F6B-805B-67AB8B4ECFF5@acm.org> ("Mattias
Engdegård"'s message of "Thu, 30 Mar 2023 12:28:30 +0200")
References: <87sfdmlgzx.fsf@gmail.com>
<87A3D742-5B8B-4F6B-805B-67AB8B4ECFF5@acm.org>
X-Draft-From: ("nnimap+imap.mail.yahoo.com:Inbox" 259379)
Date: Thu, 30 Mar 2023 19:49:42 +0800
Message-ID: <877cuyv4q1.fsf@yahoo.com>
User-Agent: Gnus/5.13 (Gnus v5.13)
P.S. here's another example of what I mean, on x86_64:
int
alignment_test (char *c)
{
register long i, *x;
x = (long *) c;
for (i = 0; i < 64; ++i)
x[i] += 1;
}
with `gcc7 -O3', this generates:
alignment_test:
movq %rdi, %rax <------- x = x1
xorl %ecx, %ecx
shrq $3, %rax <--------- x1 /= 8
andl $1, %eax <--------- x1 &= 1
je .L2 <--------- if so, assume it is already 16 byte aligned
addq $1, (%rdi) <------- otherwise, it is only 8 byte aligned. Add
1 to the first long.
movl $1, %ecx
.L2:
movl $64, %r9d
movdqa .LC0(%rip), %xmm1 <----------- load vector of 1s
subq %rax, %r9
leaq (%rdi,%rax,8), %rsi <--------- x2 = x + (x1 * 8)
xorl %edx, %edx
movq %r9, %r8
xorl %eax, %eax <--------- i = 0
shrq %r8
.L3:
movdqa (%rsi,%rax), %xmm0 <--------- Load 16 byte aligned x2[i]!!
addq $1, %rdx
paddq %xmm1, %xmm0
movaps %xmm0, (%rsi,%rax) <--------- Store 16 byte aligned x2[i]!!
addq $16, %rax <--------- Next 16 bytes...
cmpq %r8, %rdx
jb .L3
movq %r9, %rdx
andq $-2, %rdx
cmpq %rdx, %r9
leaq (%rdx,%rcx), %rax
je .L4
addq $1, (%rdi,%rax,8)
.L4:
rep ret
.LC0:
.quad 1
.quad 1
movqda traps if its first operand is not aligned to 16 bytes. When
`alignment_test' is called without an 8 byte aligned argument, that is
what will happen.
--- End Message ---