[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH v11 23/29] target/arm: [tcg] Port to translate_i

From: Richard Henderson
Subject: Re: [Qemu-devel] [PATCH v11 23/29] target/arm: [tcg] Port to translate_insn
Date: Fri, 7 Jul 2017 07:48:37 -1000
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.2.1

On 07/07/2017 07:33 AM, Peter Maydell wrote:
On 7 July 2017 at 16:26, Richard Henderson <address@hidden> wrote:
That is the purpose of DISAS_EXC too.  We've called a noreturn helper to
raise an exception and all following code is unreached.  If there *was* any
code being emitted afterward, that is arguably a bug.

One exception to that is a conditionally executed
exception generating exception -- there will in that
case be a following label for the condfail case to branch
to and the code for the condfail path.

The distinction in the case that this code fragment is touching
is that the cases handled in current master via 'goto
done_generating' and in Lluis' patch as
DISAS_SKIP are the "this insn is going to generate an
exception without even thinking about conditional
exception" (ie breakpoints, singlestep); DISAS_EXC
is for "the instruction itself generates an exception,
so don't bother with emitting too much unreachable
code to update the PC etc, but we still need to handle
the usual end-of-insn condfail path".


LLuis, this implies that the DISAS_NORETURN that I talked about elsewhere should be the thing handled generically, but that target/arm still needs a target-specific define for DISAS_EXC so that the conditional execution handler can make the distinction.

We do a few things in the DISAS_EXC codepath
(like calling gen_set_condexec()) which strictly speaking
are pointless but which it didn't seem worth trying to
avoid just to avoid generating a few extra bytes in the
generated code in a not-terribly-likely case.

Yeah. We'd probably be better off just adding dead-code removal to TCG. Something that used to be difficult but would now be trivial to do.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]