Re: [Qemu-devel] [PATCH v1 1/2] tcg: Add support for constant value promises

From: "Lluís Vilanova" <vilanova@ac.upc.edu>
To: Richard Henderson <rth@twiddle.net>
Cc: "Edgar E . Iglesias" <edgar.iglesias@gmail.com>,
	qemu-devel@nongnu.org, peter.maydell@linaro.org
Subject: Re: [Qemu-devel] [PATCH v1 1/2] tcg: Add support for constant value promises
Date: Fri, 15 Jan 2016 21:12:09 +0100	[thread overview]
Message-ID: <87egdi1u1y.fsf@fimbulvetr.bsc.es> (raw)
In-Reply-To: <56993854.1030802@twiddle.net> (Richard Henderson's message of "Fri, 15 Jan 2016 10:20:04 -0800")

Richard Henderson writes:

> On 01/15/2016 07:35 AM, Lluís Vilanova wrote:
>> +TCGv_i64 tcg_promise_i64(TCGv_promise_i64 *promise)
>> +{
>> +    int pi = tcg_ctx.gen_next_parm_idx;
>> +    *promise = (TCGv_promise_i64)&tcg_ctx.gen_opparam_buf[pi];
>> +    return tcg_const_i64(0xdeadcafe);
>> +}

> This doesn't work for a 32-bit host.  The constant may be split across two
> different parameter indices, and you don't know exactly where the second will be.

> Because of that, I think this is over-engineered, and really prefer the simpler
> interface that Edgar posted last week.

In this case, 'tcg_set_promise_i64' sets the two arguments accordingly on 32-bit
targets. Both solutions depend on TCG internals (in this specific case the
implementation of 'tcg_gen_movi_i64'), but now it's all implemented inside TCG.

Alternatively, promises could use the longer route of recording the opcode index
(as Edgar did AFAIR), and retrieve the argument pointer from there. Still, for
32-bit targets we have to assume the two immediate moves are gonna generate two
consecutive opcodes.

Thanks,
  Lluis