[PATCH] Splitting of virtual definitions (aka. V_MUST_DEF)

[Steven Bosscher]

On Tuesday 08 June 2004 06:01, Geert Bosch wrote:
> On Jun 7, 2004, at 23:08, Jeffrey A Law wrote:
> > Having a V_MUST_DEF allows us to do some interest things as it
> > indicates a more precise level of aliasing information.  ie, it
> > allows us to say that a highly specific area of memory is going
> > to be changed.  With that information we can improve redundant
> > load/store elimination in addition to improving alias analysis.
>
> I understand, but when we add 3% to compile time as a result
> of this,

Hi Geert,

First, your math is interesting, but clearly wrong:

Before: 37m37s  -> 2257s
After:  37m57s  -> 2277s
2277/2257 = 1.009, according to my calculator.

So your "3%" is in fact less than 1%, which could be
just noise.

But I think those numbers are pretty irrelevant anyway
since Brian doesn't state if they're averages of, say,
3 runs; and with a system time of more than 6 minutes
I think (well, hope) this was tested on a machine that
was also crunching cycles on something else.


> we should have a double look at the data structures
> we use. How do other compilers record this data?

Andrew Macleod already went over the data structures
we use for SSA operands and what we have now is quite
efficient.

Gr.
Steven