[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20160615195345.1af40b40@ncopa-desktop.alpinelinux.org>
Date: Wed, 15 Jun 2016 19:53:45 +0200
From: Natanael Copa <ncopa@...inelinux.org>
To: Rich Felker <dalias@...c.org>
Cc: musl@...ts.openwall.com
Subject: Re: [PATCH 2/2] refactor name_from_dns
On Fri, 27 May 2016 00:40:57 -0400
Rich Felker <dalias@...c.org> wrote:
> On Wed, May 25, 2016 at 11:22:14AM +0200, Natanael Copa wrote:
> > loop over an address family / resource record mapping to avoid
> > repetitive code.
> > ---
> > src/network/lookup_name.c | 27 +++++++++++++--------------
> > 1 file changed, 13 insertions(+), 14 deletions(-)
> >
> > diff --git a/src/network/lookup_name.c b/src/network/lookup_name.c
> > index d3d97b4..99364ab 100644
> > --- a/src/network/lookup_name.c
> > +++ b/src/network/lookup_name.c
> > @@ -141,20 +141,19 @@ static int name_from_dns(struct address buf[static MAXADDRS], char canon[static
> > int qlens[2], alens[2];
> > int i, nq = 0;
> > struct dpc_ctx ctx = { .addrs = buf, .canon = canon };
> > -
> > - if (family != AF_INET6) {
> > - qlens[nq] = __res_mkquery(0, name, 1, RR_A, 0, 0, 0,
> > - qbuf[nq], sizeof *qbuf);
> > - if (qlens[nq] == -1)
> > - return EAI_NONAME;
> > - nq++;
> > - }
> > - if (family != AF_INET) {
> > - qlens[nq] = __res_mkquery(0, name, 1, RR_AAAA, 0, 0, 0,
> > - qbuf[nq], sizeof *qbuf);
> > - if (qlens[nq] == -1)
> > - return EAI_NONAME;
> > - nq++;
> > + struct { int af; int rr; } afrr[2] = {
> > + { .af = AF_INET6, .rr = RR_A },
> > + { .af = AF_INET, .rr = RR_AAAA },
> > + };
> > +
> > + for (i=0; i<2; i++) {
> > + if (family != afrr[i].af) {
> > + qlens[nq] = __res_mkquery(0, name, 1, afrr[i].rr,
> > + 0, 0, 0, qbuf[nq], sizeof *qbuf);
> > + if (qlens[nq] == -1)
> > + return EAI_NONAME;
> > + nq++;
> > + }
>
> Can you compare the codegen for this as written vs making the table
> static-const? I wonder which gcc does better with. My guess would be
> static-const is better if it actually emits the table and loops over
> it, but auto storage like you did might be better if that helps it
> just unroll it plugging in the values.
as shown above (without static const) gcc will make it loop (there is
only one call to __res_mkquery):
name_from_dns:
pushq %r15
pushq %r14
movq %rdi, %r9
pushq %r13
pushq %r12
movl %ecx, %r11d
pushq %rbp
pushq %rbx
movl $6, %ecx
movq %r8, %r10
xorl %r12d, %r12d
xorl %ebp, %ebp
subq $1720, %rsp
leaq 128(%rsp), %r13
leaq 688(%rsp), %rbx
leaq 104(%rsp), %rdi
leaq 88(%rsp), %r14
movq %rdx, (%rsp)
movl $10, 88(%rsp)
leaq 280(%r13), %rax
movq %r13, 56(%rsp)
movq %rbx, 72(%rsp)
movl $1, 92(%rsp)
movl $2, 96(%rsp)
movq %rax, 64(%rsp)
leaq 512(%rbx), %rax
movl $28, 100(%rsp)
movq %rax, 80(%rsp)
xorl %eax, %eax
rep stosl
leaq 104(%rsp), %rax
movq %r9, 104(%rsp)
movq %rsi, 112(%rsp)
movq %rax, 8(%rsp)
.L17:
cmpl %r11d, (%r12,%r14)
je .L14
movslq %ebp, %r15
movq %r10, 24(%rsp)
movl %r11d, 20(%rsp)
imulq $280, %r15, %rax
pushq %rsi
pushq $280
movl 4(%r14,%r12), %ecx
xorl %r9d, %r9d
xorl %r8d, %r8d
xorl %edi, %edi
movl $1, %edx
addq %r13, %rax
pushq %rax
pushq $0
movq 32(%rsp), %rsi
call __res_mkquery
movl %eax, 72(%rsp,%r15,4)
addq $32, %rsp
incl %eax
movl 20(%rsp), %r11d
movq 24(%rsp), %r10
jne .L15
.L20:
movl $-2, %edx
jmp .L16
.L15:
incl %ebp
.L14:
addq $8, %r12
cmpq $16, %r12
jne .L17
leaq 48(%rsp), %r12
leaq 72(%rsp), %rcx
leaq 40(%rsp), %rdx
leaq 56(%rsp), %rsi
pushq %rax
pushq %r10
movl $512, %r9d
movq %r12, %r8
movl %ebp, %edi
call __res_msend_rc
xorl %r13d, %r13d
testl %eax, %eax
popq %rdx
movl $-11, %edx
popq %rcx
js .L16
.L18:
cmpl %r13d, %ebp
jle .L31
movl (%r12,%r13,4), %esi
movq 8(%rsp), %rcx
leaq dns_parse_callback(%rip), %rdx
movq %rbx, %rdi
incq %r13
addq $512, %rbx
call __dns_parse
jmp .L18
.L31:
movl 120(%rsp), %edx
testl %edx, %edx
jne .L16
cmpl $3, 48(%rsp)
jle .L23
movb 691(%rsp), %al
andl $15, %eax
cmpb $2, %al
je .L23
testb %al, %al
je .L20
cmpb $3, %al
movl $-4, %eax
cmovne %eax, %edx
jmp .L16
.L23:
...
with "static const struct { int af; int rr; } afrr[2] = {" gcc will
actually unroll it (there are 2 calls to __res_mkquery):
name_from_dns:
pushq %r15
pushq %r14
movq %rdi, %r9
pushq %r13
pushq %r12
movl %ecx, %r15d
pushq %rbp
pushq %rbx
movl $6, %ecx
movq %rdx, %r13
movq %r8, %r14
subq $1688, %rsp
leaq 96(%rsp), %r12
leaq 656(%rsp), %rbx
leaq 72(%rsp), %rdi
leaq 280(%r12), %rax
movq %r12, 40(%rsp)
movq %rbx, 56(%rsp)
movq %rax, 48(%rsp)
leaq 512(%rbx), %rax
movq %rax, 64(%rsp)
xorl %eax, %eax
cmpl $10, %r15d
rep stosl
leaq 72(%rsp), %rax
movq %r9, 72(%rsp)
movq %rsi, 80(%rsp)
movq %rax, 8(%rsp)
je .L21
pushq %rdi
pushq $280
xorl %r9d, %r9d
pushq %r12
pushq $0
xorl %r8d, %r8d
xorl %edi, %edi
movl $1, %ecx
movl $1, %edx
movq %r13, %rsi
call __res_mkquery
movl %eax, 56(%rsp)
addq $32, %rsp
incl %eax
je .L15
cmpl $2, %r15d
movl $1, %ebp
jne .L36
jmp .L16
.L21:
xorl %ebp, %ebp
.L36:
movslq %ebp, %r15
pushq %rsi
pushq $280
imulq $280, %r15, %rax
xorl %r9d, %r9d
xorl %r8d, %r8d
xorl %edi, %edi
movl $28, %ecx
movl $1, %edx
movq %r13, %rsi
incl %ebp
addq %rax, %r12
pushq %r12
pushq $0
call __res_mkquery
movl %eax, 56(%rsp,%r15,4)
addq $32, %rsp
incl %eax
jne .L16
.L15:
movl $-2, %edx
jmp .L18
.L16:
leaq 32(%rsp), %r12
leaq 56(%rsp), %rcx
leaq 24(%rsp), %rdx
leaq 40(%rsp), %rsi
pushq %rax
pushq %r14
movl $512, %r9d
movq %r12, %r8
movl %ebp, %edi
call __res_msend_rc
xorl %r14d, %r14d
testl %eax, %eax
popq %rdx
movl $-11, %edx
popq %rcx
js .L18
.L19:
cmpl %r14d, %ebp
jle .L38
movl (%r12,%r14,4), %esi
movq 8(%rsp), %rcx
leaq dns_parse_callback(%rip), %rdx
movq %rbx, %rdi
incq %r14
addq $512, %rbx
call __dns_parse
jmp .L19
.L38:
movl 88(%rsp), %edx
testl %edx, %edx
jne .L18
cmpl $3, 32(%rsp)
jle .L24
movb 659(%rsp), %al
andl $15, %eax
cmpb $2, %al
je .L24
testb %al, %al
je .L15
cmpb $3, %al
movl $-4, %eax
cmovne %eax, %edx
jmp .L18
.L24:
>
> Rich
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
