Re: [lno] [RFC] if-conversion and auto vectorizer

[Devang Patel <dpatel at apple dot com>]

On Mar 15, 2004, at 3:20 PM, Richard Henderson wrote:

> > Concerns we have is: Do we need separate pass (which I am currently
> > implementing) to transform if-then-else to  use 'conditional modify
> > expr' for vectorizer or not?
>
> I suspect that your vectorizer will want something much stronger than
> what we will ever get from phiopts.  Phiopts is simply trying to tidy
> up simple situations.

Thank you. This helps to clarify things.

> --
>
> Something that you should watch is that COND_EXPR is *not* predication,
> and you can't force it to be predication on most targets.  Depending on
> what papers you're reading, this may significantly affect what you're
> planning on doing.
>
> The big distinction involves trapping expressions and operands.  E.g.
>
> 	if (test)
> 	   T_1 = *p;
> 	else
> 	   T_2 = 0;
> 	T_3 = PHI (T_1, T_2);
>
> cannot be transformed to
>
> 	T_1 = *p;
> 	T_3 = test ? T_1 : 0;
> unless you can show either that (1) dereferencing *p will never trap,
> or (2) some other dereference of *p dominates this test.  Similar
> examples can be shown with arithmetic such as divide, or fp arithmetic
> without -ffast-math (i.e. when we can't discount NaNs).

We also need to consider possible side effects of 'test' itself.

--
Devang