Re: [PATCH v3 17/35] x86/pci/xen: Use msi_msg shadow structs

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1.1: Type: text/plain, Size: 4547 bytes --]

On Sun, 2020-10-25 at 09:49 +0000, David Laight wrote:
> Just looking at a random one of these patches...
> 
> Does the compiler manage to optimise that reasonably?
> Or does it generate a lot of shifts and masks as each
> bitfield is set?
> 
> The code generation for bitfields is often a lot worse
> that that for |= setting bits in a word.

Indeed, it appears to be utterly appalling. That was one of my
motivations for doing it with masks and shifts in the first place.

Because in ioapic_setup_msg_from_msi(), for example, these two are
consecutive in both source and destination:

	entry->vector			= msg.arch_data.vector;
	entry->delivery_mode		= msg.arch_data.delivery_mode;

And so are those:

	entry->ir_format		= msg.arch_addr_lo.dmar_format;
	entry->ir_index_0_14		= msg.arch_addr_lo.dmar_index_0_14;

But GCC 7.5.0 here is doing them all separately...

00000000000011e0 <ioapic_setup_msg_from_msi>:
{
    11e0:       53                      push   %rbx
    11e1:       48 89 f3                mov    %rsi,%rbx
    11e4:       48 83 ec 18             sub    $0x18,%rsp
        irq_chip_compose_msi_msg(irq_data, &msg);
    11e8:       48 89 e6                mov    %rsp,%rsi
{
    11eb:       65 48 8b 04 25 28 00    mov    %gs:0x28,%rax
    11f2:       00 00 
    11f4:       48 89 44 24 10          mov    %rax,0x10(%rsp)
    11f9:       31 c0                   xor    %eax,%eax
        irq_chip_compose_msi_msg(irq_data, &msg);
    11fb:       e8 00 00 00 00          callq  1200 <ioapic_setup_msg_from_msi+0x20>
        entry->vector                   = msg.arch_data.vector;
    1200:       0f b6 44 24 08          movzbl 0x8(%rsp),%eax
        entry->delivery_mode            = msg.arch_data.delivery_mode;
    1205:       0f b6 53 01             movzbl 0x1(%rbx),%edx
    1209:       0f b6 74 24 09          movzbl 0x9(%rsp),%esi
        entry->vector                   = msg.arch_data.vector;
    120e:       88 03                   mov    %al,(%rbx)
        entry->dest_mode_logical        = msg.arch_addr_lo.dest_mode_logical;
    1210:       0f b6 04 24             movzbl (%rsp),%eax
        entry->delivery_mode            = msg.arch_data.delivery_mode;
    1214:       83 e2 f0                and    $0xfffffff0,%edx
    1217:       83 e6 07                and    $0x7,%esi
        entry->dest_mode_logical        = msg.arch_addr_lo.dest_mode_logical;
    121a:       09 f2                   or     %esi,%edx
    121c:       8d 0c 00                lea    (%rax,%rax,1),%ecx
        entry->ir_format                = msg.arch_addr_lo.dmar_format;
    121f:       c0 e8 04                shr    $0x4,%al
        entry->dest_mode_logical        = msg.arch_addr_lo.dest_mode_logical;
    1222:       83 e1 08                and    $0x8,%ecx
    1225:       09 ca                   or     %ecx,%edx
    1227:       88 53 01                mov    %dl,0x1(%rbx)
        entry->ir_format                = msg.arch_addr_lo.dmar_format;
    122a:       89 c2                   mov    %eax,%edx
        entry->ir_index_0_14            = msg.arch_addr_lo.dmar_index_0_14;
    122c:       8b 04 24                mov    (%rsp),%eax
    122f:       83 e2 01                and    $0x1,%edx
    1232:       c1 e8 05                shr    $0x5,%eax
    1235:       66 25 ff 7f             and    $0x7fff,%ax
    1239:       8d 0c 00                lea    (%rax,%rax,1),%ecx
    123c:       66 c1 e8 07             shr    $0x7,%ax
    1240:       88 43 07                mov    %al,0x7(%rbx)
    1243:       09 ca                   or     %ecx,%edx
}
    1245:       48 8b 44 24 10          mov    0x10(%rsp),%rax
    124a:       65 48 33 04 25 28 00    xor    %gs:0x28,%rax
    1251:       00 00 
        entry->ir_index_0_14            = msg.arch_addr_lo.dmar_index_0_14;
    1253:       88 53 06                mov    %dl,0x6(%rbx)
}
    1256:       75 06                   jne    125e <ioapic_setup_msg_from_msi+0x7e>
    1258:       48 83 c4 18             add    $0x18,%rsp
    125c:       5b                      pop    %rbx
    125d:       c3                      retq   
    125e:       e8 00 00 00 00          callq  1263 <ioapic_setup_msg_from_msi+0x83>
    1263:       0f 1f 00                nopl   (%rax)
    1266:       66 2e 0f 1f 84 00 00    nopw   %cs:0x0(%rax,%rax,1)
    126d:       00 00 00 


Still, this isn't really a fast path so I'm content to file the GCC PR
for the really *obvious* misses and let it take its course.

[-- Attachment #1.2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5174 bytes --]

[-- Attachment #2: Type: text/plain, Size: 156 bytes --]

_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu