Re: Question on implementation detail of `temp_sync`

* Re: Question on implementation detail of `temp_sync`
       [not found] ` <tencent_3C5D583315945B14647C946B@qq.com>
@ 2020-08-05  8:39   ` Alex Bennée
  0 siblings, 0 replies; 2+ messages in thread
From: Alex Bennée @ 2020-08-05  8:39 UTC (permalink / raw)
  To: lrwei; +Cc: Richard Henderson, QEMU Developers, qemu-discuss

lrwei <lrwei@bupt.edu.cn> writes:

> Sorry for the unintentional sending of an uncompleted message.

Questions about the internals of the TCG are very much in the remit of
qemu-devel so are likely to get missed on qemu-discuss which is more
aimed at user questions.

>
<re-pasted to fix html noise>

> I understands that the current code works, but gets confused on why `ts` needs to be loaded in to a register when `free_or_dead` is not
> set.

It isn't, the break leaves the switch statement once it stores the
constant to memory.

> For example in the following scenario:
> movi_i32    r0, 0x1
> add_i32      r1, r1, r0
> ...
> (where r0 is not used any more, and both r0 and r1 are globals)

> If I am not mistaken, the code gen procedure of the first IR will call `temp_sync` with `free_or_dead` not set, which load the constant in to
> a register and store it back to memory. At this time, `r0` will be `TEMP_VAL_REG` instead of `TEMP_VAL_CONST`, so the following IR can't
> embed this constant operand in the assembly instruction it produces. Also, this results in a seemingly useless register allocation (, why
> don't the further use of r0 use the constant directly?)

Is this what you are actually seeing generated? If you run with -d
in_asm,op,op_opt,out_asm it should be clear what actually happened.

> So I wonder whether there is any reason for this loading a constant into register, I'll be very appreciated if someone can point out the
> reason for me.

<snip>
>
>
> Thanks in advance.
> lrwei&nbsp; 
> &nbsp;
> ------------------&nbsp;Original&nbsp;------------------
> From: &nbsp;"lrwei"<lrwei@bupt.edu.cn&gt;;
> Date: &nbsp;Tue, Aug 4, 2020 12:06 PM
> To: &nbsp;"qemu-discuss"<qemu-discuss@nongnu.org&gt;; 
> Subject: &nbsp;Question on implementation detail of `temp_sync`
>
<re-pasted fixing html noise>

> Hello to the list,
> Recently I have been studying the code of TCG, and get confused by the following detail in function `temp_sync` in tcg/tcg.c:

>     case TEMP_VAL_CONST:
>         /* If we're going to free the temp immediately, then we won't
>            require it later in a register, so attempt to store the
>            constant to memory directly.  */
>         if (free_or_dead
>            && tcg_out_sti(s, ts->type, ts->val,
>                            ts->mem_base->reg, ts->mem_offset)) {
>            break;
>         }
>         temp_load(s, ts, tcg_target_available_regs[ts->type],
>                   allocated_regs, preferred_regs);
>         /* fallthrough */

> movi_i32

> Would it be better to remove the `free_or_dead` in the if statement, i.e. turn the function to be:

>     case TEMP_VAL_CONST:
>         if (tcg_out_sti(s, ts->type, ts->val,
>                            ts->mem_base->reg, ts->mem_offset)) {
>            break;
>         }
>         temp_load(s, ts, tcg_target_available_regs[ts->type],
>                   allocated_regs, preferred_regs);
>         /* fallthrough */

-- 
Alex Bennée

^ permalink raw reply	[flat|nested] 2+ messages in thread