Re: [PATCH 5.15 00/78] 5.15.55-rc1 review

[Paolo Bonzini]

On 7/14/22 19:02, Linus Torvalds wrote:
> And guess what? The code could just use roundup_pow_of_two(), which is
> designed exactly like ilog2() to be used for compile-time constant
> values.
>
> So the code should just use
>
>      #define FASTOP_SIZE roundup_pow_of_two(FASTOP_LENGTH)
>
> and be a lot more legible, wouldn't it?
>
> Because I don't think there is anything magical about the length
> "8/16/32". It's purely "aligned and big enough to contain
> FASTOP_LENGTH".

roundup_pow_of_two unfortunately is not enough for stringizing 
FASTOP_SIZE into an asm statement. :(

	#define __FOP_FUNC(name) \
	        ".align " __stringify(FASTOP_SIZE) " \n\t" \
	        ".type " name ", @function \n\t" \
	        name ":\n\t" \
	        ASM_ENDBR

The shifts are what we came up with for the SETCC thunks when ENDBR and 
SLS made them grew beyond 4 bytes; Peter's patch is reusing the trick 
for the fastop thunks.

> Because I don't think there is anything magical about the length
> "8/16/32". It's purely "aligned and big enough to contain
> FASTOP_LENGTH".

I agree with that, it's only limited to 8/16/32 to keep the macro to a 
decent size.

> And then the point of that
>
>     static_assert(FASTOP_LENGTH <= FASTOP_SIZE);
>
> just goes away, because there are no subtle math issues there any more.
>
> In fact, the remaining question is just "where did the 7 come from" in
>
>     #define FASTOP_LENGTH (7 + ENDBR_INSN_SIZE + RET_LENGTH)

The 7 is an upper limit to the length of the code between endbr and ret.
There's no particular reason to limit to 7, but it allows using an 
alignment of 8 in the smallest case (no thunks, no SLS, no endbr) where 
you just have ".align 8; ...; ret".

Paolo