Re: [PATCH 1/4] util/cacheflush: fix illegal instruction on windows-arm64

[Pierrick Bouvier]

After some investigation on this, I found that even faking ctr_el0 
content does not work. Indeed, "dc cvau" and "ic ivau" both are 
privileged too (and fail with illegal instruction).

I started looking how other projects are handling this. In the case of 
firefox js engine, they simply perform a FlushInstructionCache [1], and 
nothing for data cache. From what I found on various websites, there is 
no way to perform any data cache flush from userspace under Windows.

Out of curiosity, I looked in llvm how __builtin___clear_cache was 
implemented, and for windows-arm64, it's simply a call to the same 
function, FlushInstructionCache [2].

This explains why the generic implementation of flush_idcache_range does 
the correct thing for this platform, and why tests I ran worked. Thus, 
it's probably the best approach we can have for now.

[1] 
https://hg.mozilla.org/mozilla-central/file/tip/js/src/jit/arm64/vixl/MozCpu-vixl.cpp#l110
[2] 
https://github.com/llvm/llvm-project/blob/574e417460cdfebb8157fbe61b6a015e44856f64/compiler-rt/lib/builtins/clear_cache.c#L66

On 2/15/23 19:22, Richard Henderson wrote:
> On 2/15/23 02:49, Pierrick Bouvier wrote:
> > I'm not an expert on this area, but I imagine that booting a full VM will force 
> > TCG to
> > emit code at the same address twice (after having generated enough translated 
> > blocks),
> > which shows that generic flush_idcache_range works. Is that reasoning correct?
> >
> > Do you think we still need a specialized version for windows-arm64?
>
> No, I don't think so.  It would be ideal if we were able to read CTR_EL0, 
> because we make
> use of the IDC field, but 0 is a safe default.
>
>
> r~