[Qemu-devel] TCG and branches

[Qemu-devel] TCG and branches

[Benjamin Herrenschmidt]

Hi Richard !

So in my discovery of TCG, one thing I noticed is the horrendous amount
of code generated for branches, especially conditional ones.

I have a patch at least to remove a bunch of dead gunk on target-ppc
for non-conditional ones (we still generated the "else" of the
condition even when never branching to it).

However, I wonder if there are ways to do better.

The first obvious thing that comes to mind is to avoid stopping the
TB on a non-taken conditional branch. The reason that can't be done
today, from my limited understanding of things, is because we only
support index 0 and 1 today for gen_goto_tb().

Now I haven't completely figured out how TB linkage works under the
hood but do you know of any fundamental reason why we have that limit ?

Could we, for example, have a limit of, for example, 8 and only break
the TB after a branch if we have less than 2 left ?

Or are there deeper reasons why we really can't link more than 2 ?

With the current implementation, a non-taken branch results in three
branches: One to "skip over" the condition-true case, one to the new TB
and one in the header of that new TB to check for exits. We could
reduce to one in most cases.

Cheers,
Ben.

Re: [Qemu-devel] [Qemu-ppc] TCG and branches

[Nikunj A Dadhania]

Benjamin Herrenschmidt <benh@kernel.crashing.org> writes:

> Hi Richard !
>
> So in my discovery of TCG, one thing I noticed is the horrendous amount
> of code generated for branches, especially conditional ones.
>

static inline void gen_bcond(DisasContext *ctx, int type)
{
[...]
    if ((bo & 0x4) == 0) {
        /* Decrement and test CTR */
        [...]
        if (bo & 0x2) {
            tcg_gen_brcondi_tl(TCG_COND_NE, temp, 0, l1);
        } else {
            tcg_gen_brcondi_tl(TCG_COND_EQ, temp, 0, l1);
        }

BUG, both jumping to same label?

Can we replace it unconditionally with

tcg_gen_br(l1) ?

Regards
Nikunj

Re: [Qemu-devel] [Qemu-ppc] TCG and branches

[Nikunj A Dadhania]

Nikunj A Dadhania <nikunj@linux.vnet.ibm.com> writes:

> Benjamin Herrenschmidt <benh@kernel.crashing.org> writes:
>
>> Hi Richard !
>>
>> So in my discovery of TCG, one thing I noticed is the horrendous amount
>> of code generated for branches, especially conditional ones.
>>
>
> static inline void gen_bcond(DisasContext *ctx, int type)
> {
> [...]
>     if ((bo & 0x4) == 0) {
>         /* Decrement and test CTR */
>         [...]
>         if (bo & 0x2) {
>             tcg_gen_brcondi_tl(TCG_COND_NE, temp, 0, l1);
>         } else {
>             tcg_gen_brcondi_tl(TCG_COND_EQ, temp, 0, l1);
>         }
>
> BUG, both jumping to same label?

Sorry for the noise, I understood it wrong.

Regards
Nikunj

Re: [Qemu-devel] TCG and branches

[Richard Henderson]

On 07/31/2016 05:23 AM, Benjamin Herrenschmidt wrote:
> The first obvious thing that comes to mind is to avoid stopping the
> TB on a non-taken conditional branch. The reason that can't be done
> today, from my limited understanding of things, is because we only
> support index 0 and 1 today for gen_goto_tb().

Correct.

> Now I haven't completely figured out how TB linkage works under the
> hood but do you know of any fundamental reason why we have that limit ?
>
> Could we, for example, have a limit of, for example, 8 and only break
> the TB after a branch if we have less than 2 left ?

We do borrow low bits of a pointer (to a TranslationBlock) in implementing 
exit_tb.  See TB_EXIT_MASK and the large block comment just above its 
definition in tcg/tcg.h.

However, increasing the alignment of TranslationBlock ought to be trivial, 
giving you as many bits as required.  With that, I see no reason in principal 
that this wouldn't work.

Such a change would need some measurement to see how often this occurs, and how 
much this helps the actual runtime performance.


r~