On Tue, Apr 28, 2020 at 12:43:54PM +0100, Catalin Marinas wrote:
> Hi Dave,
>
> On Mon, Apr 27, 2020 at 05:57:37PM +0100, Dave P Martin wrote:
> > On Tue, Apr 21, 2020 at 03:25:41PM +0100, Catalin Marinas wrote:
> > > There are situations where we do not want to disable the whole block
> > > based on a config option, only the alternative part while keeping the
> > > first instruction. Improve the alternative_insn assembler macro to take
> > > a 'first_insn' argument, default 0, to preserve the current behaviour.
> > >
> > > Signed-off-by: Catalin Marinas <catalin.marinas@xxxxxxx>
> > > Cc: Will Deacon <will@xxxxxxxxxx>
> > > ---
> > > arch/arm64/include/asm/alternative.h | 8 +++++++-
> > > 1 file changed, 7 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/arch/arm64/include/asm/alternative.h b/arch/arm64/include/asm/alternative.h
> > > index 5e5dc05d63a0..67d7cc608336 100644
> > > --- a/arch/arm64/include/asm/alternative.h
> > > +++ b/arch/arm64/include/asm/alternative.h
> > > @@ -111,7 +111,11 @@ static inline void apply_alternatives_module(void *start, size_t length) { }
> > > .byte \alt_len
> > > .endm
> > >
> > > -.macro alternative_insn insn1, insn2, cap, enable = 1
> > > +/*
> > > + * Disable the whole block if enable == 0, unless first_insn == 1 in which
> > > + * case insn1 will always be issued but without an alternative insn2.
> > > + */
> > > +.macro alternative_insn insn1, insn2, cap, enable = 1, first_insn = 0
> > > .if \enable
> > > 661: \insn1
> > > 662: .pushsection .altinstructions, "a"
> > > @@ -122,6 +126,8 @@ static inline void apply_alternatives_module(void *start, size_t length) { }
> > > 664: .popsection
> > > .org . - (664b-663b) + (662b-661b)
> > > .org . - (662b-661b) + (664b-663b)
> > > + .elseif \first_insn
> > > + \insn1
> >
> > This becomes quite unreadable at the invocation site, especially when
> > invoked as "alternative_insn ..., 1". "... first_insn=1" is not much
> > better either).
>
> That I agree.
>
> The reason I didn't leave the alternative in place here is that if gas
> doesn't support MTE, it will fail to compile. I wanted to avoid the
> several #ifdef's.
We could solve that by synthesising the opcodes instead of relying on
gas (as we do for other extensions).
But I'd agree that's just pushing the problem around rather than solving
it. It seems dumb to go to that trouble for a case where the affected
insn isn't going to be emitted...
> > I'm struggling to find non-trivial users of this that actually want the
> > whole block to be deleted dependent on the config.
>
> Some of the errata stuff like CONFIG_ARM64_REPEAT_TLBI ends up with
> unnecessary nops. Similarly for CONFIG_ARM64_UAO/PAN and maybe a few
> others (it's all additional nops). We also have a few errata workaround
> where we didn't bother with the config enable option at all.
OK, looks like I may have missed some cases. There's a dense thicket of
macros that call each other here, and I've not looked at it for a while ;)
> While this is C code + inline asm, I'd like to have a consistent
> behaviour of ALTERNATIVE between C and .S files. Now, given that some of
> them (like UAO/PAN) are on by default, it probably doesn't make any
> difference if we always keep the first block (non-alternative).
>
> We could add a new macro ALTERNATIVE_OR_NOP.
alternative_insn doesn't seem exist for C at all. Did I miss something?
> > Can we instead just always behave as if first_insn=1 instead? This this
> > works intuitively as an alternative, not the current weird 3-way choice
> > between insn1, insn2 and nothing at all. The only time that makes sense
> > is when one of the insns is a branch that skips the block, but that's
> > handled via the alternative_if macros instead.
> >
> > Behaving always like first_insn=1 provides an if-else that is statically
> > optimised if the relevant feature is configured out, which I think is
> > the only thing people are ever going to want.
> >
> > Maybe something depends on the current behaviour, but I can't see it so
> > far...
>
> I'll give it a go in v4 and see how it looks.
>
> Another option would be an alternative_else which takes an enable
> argument.
Sure, I think it could make sense to have a different wrapper so that
the meaning of invocations is clearer for this special case.
For the underlying macro, maybe it would be simpler to make it truly
3-way:
.macro alternative_insn insn_with_cap:req, insn_without_cap:req, cap:req, \
enable_alternative=1, fallback_insn=
// ...
.if (\enable_alternative)
// as currently
.else
\fallback_insn
.endif
.endm
Then we can rejig the various frontends around that.
If you don't want anything when the alternative is disabled, you just
omit fallback_insn.
Cheers
---Dave
