Re: [PATCH 11/25] Simplify vec_merge according to the mask.

[Andrew Stubbs <ams at codesourcery dot com>]

On 27/09/18 08:16, Richard Sandiford wrote:
> On keeping the complexity down:
>
>    if (side_effects_p (x))
>      return NULL_RTX;
>
> makes this quadratic for chains of unary operations.  Is it really
> needed?  The code after it simply recurses on operands and doesn't
> discard anything itself, so it looks like the VEC_MERGE call to
> side_effects_p would be enough.

The two calls do not check the same thing. The other one checks the 
other operand of a vec_merge, and this checks the current operand.

I suppose it's safe to discard a VEC_MERGE when the chosen operand 
contains side effects, but I'm not so sure when the VEC_MERGE itself is 
an operand to an operator with side effects. I'm having a hard time 
inventing a scenario in which a PRE_INC could contain a VEC_MERGE, but 
maybe a volatile MEM or ASM_OPERANDS could do?

Conversely, I don't see that side-effects deep down in an expression 
should stop us transforming it as a high level.

Is there an equivalent to side_effects_p that doesn't recurse? Should 
there be?

Andrew