On Fri, Nov 18, 2016 at 05:52:12PM -0500, Michael Meissner wrote:
> On Fri, Nov 18, 2016 at 04:43:40PM -0600, Segher Boessenkool wrote:
> > On Fri, Nov 18, 2016 at 03:38:38PM -0500, Michael Meissner wrote:
> > > This patch tweaks the movdi constraints for the PowerPC to use "^" or "$"
> > > constraints instead of "?*". This allows the register allocator to more
> > > often
> > > allocate DImode to a floating point/vector register when it is desirable
> > > to do
> > > so.
> >
> > It also changes some plain "?" to "^" or "$" or even "*" (which cannot
> > work for multi-character constraints, it skips one character, not one
> > constraint). Wrong version of the patch?
>
> Note, if '*' does not work with multi-character prefixes, that is a bug. All
> of rs6000.md assumes that ?*wa means that the register allocator will not
> consider VSX vector registers for when calculating register preferences.
The documentation is out of date.
>From ira-costs.c:
/* Scan all the constraint letters. See if the operand
matches any of the constraints. Collect the valid
register classes and see if this operand accepts
memory. */
while ((c = *p))
{
switch (c)
{
case '*':
/* Ignore the next letter for this pass. */
c = *++p;
break;
and then 83 lines later:
p += CONSTRAINT_LEN (c, p);
if (c == ',')
break;
}
so it does in fact work.
Neither the patch description nor the changelog says you are doing these
changes though.
> > > I built bootstrap compilers and did make check with no regressions on:
> > > 1) Little endian power8, --with-cpu=power8
> > > 2) Big endian power8, --with-cpu=power8 (no 32-bit support)
> > > 3) Big endian power7, --with-cpu=power7 (both 32/64-bit support)
> >
> > Could you also test with reload please? Just LE is enough I guess.
> > We'd like to keep reload working for GCC 7 at least, and these cost
> > prefixes tend to break mov patterns :-/
>
> Argh, I guess you are right, but then if reload doesn't work, I will likely
> submit the patch where there are three different movdi's (one for 32-bit
> without the change, one for 64-bit with reload, and one for 64-bit with lra).
> I would prefer not to do that.
Let's hope it just works :-)
Segher
