RE: [PATCH] x86: bitops: fix build regression
From: David Laight
Date: Sun May 10 2020 - 07:59:32 EST
From: Peter Anvin
> Sent: 08 May 2020 18:32
> On 2020-05-08 10:21, Nick Desaulniers wrote:
> >>
> >> One last suggestion. Add the "b" modifier to the mask operand: "orb
> >> %b1, %0". That forces the compiler to use the 8-bit register name
> >> instead of trying to deduce the width from the input.
> >
> > Ah right: https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#x86Operandmodifiers
> >
> > Looks like that works for both compilers. In that case, we can likely
> > drop the `& 0xff`, too. Let me play with that, then I'll hopefully
> > send a v3 today.
> >
>
> Good idea. I requested a while ago that they document these modifiers; they
> chose not to document them all which in some ways is good; it shows what they
> are willing to commit to indefinitely.
I thought the intention here was to explicitly do a byte access.
If the constant bit number has had a div/mod by 8 done on it then
the address can be misaligned - so you mustn't do a non-byte sized
locked access.
OTOH the original base address must be aligned.
Looking at some instruction timing, BTS/BTR aren't too bad if the
bit number is a constant. But are 6 or 7 clocks slower if it is in %cl.
Given these are locked RMW bus cycles they'll always be slow!
How about an asm multi-part alternative that uses a byte offset
and byte constant if the compiler thinks the mask is constant
or a 4-byte offset and 32bit mask if it doesn't.
The other alternative is to just use BTS/BTS and (maybe) rely on the
assembler to add in the word offset to the base address.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)