Re: [PATCH 1/1] target/i386: Raise #GP on unaligned m128 accesses when required.

[Richard Henderson]

On 8/29/22 13:46, Ricky Zhou wrote:
> Thanks for taking a look at this - did you see the bit in the cover
> letter where I discuss doing this via alignment requirements on the
> memory operation? My logic was that the memop alignment checks seem to
> be more oriented towards triggering #AC exceptions (even though this is
> not currently implemented),

I missed that in the cover.  However... implementing #AC is pretty hypothetical.  It's not 
something that I've ever seen used, and not something that anyone has asked for.

> One slightly more involved way to use alignment on the MemOp could be to
> arrange to pass the problematic MemOp to do_unaligned_access and
> helper_unaligned_{ld,st}. Then we could allow CPUs to handle
> misalignment of different MemOps differently (e.g. raise #GP/SIGSEGV for
> certain ops and #AC/SIGBUS for others). For this change to x86, we could
> maybe get away with making MO_ALIGN_16 and above trigger #GP/SIGSEGV and
> everything else trigger #AC/SIGBUS. If that's a little hacky, we could
> instead add some dedicated bits to MemOp that distinguish different
> types of unaligned accesses.

There's another related problem that actually has gotten a bug report in the past: when 
the form of the address should raise #SS instead of #GP in system mode.

My initial thought was to record information about "the" memory access in the per-insn 
unwind info, until I realized that there are insns with  multiple memory operations 
requiring different treatment.  E.g. "push (%rax)", where the read might raise #GP and the 
write might raise #SS.  So I think we'd need to encode #GP vs #SS into the mmu_idx used 
(e.g. in the lsb).

However, I don't think there are any similar situations of multiple memory types affecting 
SSE, so #AC vs #GP could in fact be encoded into the per-insn unwind info.

As for SIGBUS vs SIGSEGV for SSE and user-only, you only need implement the 
x86_cpu_ops.record_sigbus hook.  C.f. the s390x version which raises PGM_SPECIFICATION -> 
SIGILL for unaligned atomic operations.


r~