[Top][All Lists]

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

Re: [Qemu-devel] Confusion regarding temporaries with branch conditional

From: Richard Henderson
Subject: Re: [Qemu-devel] Confusion regarding temporaries with branch conditional
Date: Wed, 30 Nov 2016 09:08:11 -0800
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.4.0

On 11/29/2016 11:56 PM, Nikunj A Dadhania wrote:
> Lets bring full example here.
>     TCGv nb = tcg_temp_new();
>     tcg_gen_andi_tl(nb, cpu_gpr[rB(ctx->opcode)], 0xFF);
>     tcg_gen_brcondi_tl(TCG_COND_EQ, nb, 0, l1);
>     /* do something */
>     gen_set_access_type(ctx, ACCESS_INT);
>     EA = tcg_temp_new();
>     gen_addr_register(ctx, EA);
>     tcg_gen_qemu_ld_i64(xtl, EA, ctx->mem_idx, MO_LEQ);
>     tcg_gen_addi_tl(EA, EA, 8);
>     tcg_gen_qemu_ld_i64(xth, EA, ctx->mem_idx, MO_LEQ);
>     opc = tcg_const_i32(ctx->opcode);
>     gen_helper_lxvl(cpu_env, opc, nb);      /* <--- That uses nb */
>     tcg_temp_free_i32(opc);
>     tcg_temp_free(EA);
>     gen_set_label(l1);
>     tcg_temp_free(nb);
>> The plain temporary is only valid to the end of a basic
>> block, and brcond ends a basic block. So you can free
>> the temp after the brcond but you can't do anything
>> else with it.
> In the above case, assuming that nb is a plain temporary, case nb != 0
> worked fine (by fluke?), i.e. no branch.
> While when nb == 0, failed, i.e. branch taken to l1, and just free nb. I
> am not using "nb" in this case.

This is also a good example of why you should preferentially avoid branches
within the tcg opcode stream.

In the case of lxvl, I strongly suggest that you push *everything* into the
helper.  In particular:

 (1) Passing the full instruction opcode means you've got to re-parse.
     Why are you not passing a pointer to the XT register like other
     VSX helpers?

 (2) As I read it, this is wrong, since when NB == 0, XT is assigned 0.
     Which you are not doing, having skipped over the helper.

 (3) Use cpu_ldq_data_ra within the helper to perform the memory loads.


reply via email to

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