Re: [GIT PULL] x86 fixes

From: Andy Lutomirski
Date: Mon Jul 02 2018 - 14:48:28 EST

On Sat, Jun 30, 2018 at 12:01 PM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Sat, Jun 30, 2018 at 1:49 AM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>> --- a/arch/x86/entry/entry_32.S
>> +++ b/arch/x86/entry/entry_32.S
>> @@ -477,7 +477,7 @@ ENTRY(entry_SYSENTER_32)
>> * whereas POPF does not.)
>> */
>> addl $PT_EFLAGS-PT_DS, %esp /* point esp at pt_regs->flags */
>> - btr $X86_EFLAGS_IF_BIT, (%esp)
>> + btrl $X86_EFLAGS_IF_BIT, (%esp)
>> popfl
> Ho humm. Just looking at this patch, my reaction was "why isn't this
> an 'andl $~X86_EFLAGS_IF' instead"?
> Yeah, I guess the 'andl' is two bytes longer (due to the 32-bit
> constant - because IF is bit 9, you can't use a byte constant, and you
> don't want to get a partial word write just before the popfl).
> But btr is really pretty heavy operation for older CPU's (it's gotten
> better, but 32-bit code presumably cares more about the older CPUs).
> It really doesn't matter, I guess. The btr goes back to commit
> c2c9b52fab0d ("x86/entry/32: Restore FLAGS on SYSEXIT").
> Andy?

BTR is way more leet than AND!

Seriously, though, I've never really tried to shave cycles off the
32-bit code, and BTR is shorter, and I didn't spend more than about
one brain cycle thinking about it. I guess that BTR has a more
complicated flags pipeline (the output flags depend on the input, not
just the output) and probably uses some more complicated ALU circuit
as compared to ANDL.