This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: patch to enable LRA for ppc


On Mon, Oct 21, 2013 at 08:18:22PM +0200, Steven Bosscher wrote:
> On Mon, Oct 21, 2013 at 5:51 PM, Michael Meissner wrote:
> > What I'm doing is adding secondary reload support so that up until reload time,
> > we can represent VSX addresses as reg+offset, and in secondary reload, create
> > the addition instructions to put the offset in a base register.  I haven't made
> > any changes to the machine independent portions of the compiler.  As long as
> > IRA uses the secondary reload interface, it should be ok.  However, right now,
> > I need to focus most of my attention on getting the secondary reload support to
> > work.
> >
> > One thing that I've asked for before, but to remind you, is I really, really
> > wish secondary reload could allocate two scratch registers if it is given an
> > insn that takes 4 arguments.  Right now, I'm allocating a TFmode scratch, since
> > that gives 2 registers, but future changes will want TFmode to go into a single
> > vector register, and I will need to create another type, like V4DI that does
> > take 2 registers.  The case that this is needed for is moving an item from GPRs
> > to VSX registers that takes 2 GPR registers, such as moving 128-bit items in
> > 64-bit mode, or 64-bit items in 32-bit mode.  I need two registers to do the
> > move into, and then I will do the combine operation.
> 
> 
> Eh, perhaps I'm missing something, but...
> 
> Isn't one of the great advantages of LRA over reload, that LRA allows
> you to create new pseudos so that you shouldn't ever need secondary
> reloads??

You still need secondary reload.

For example, on the powerpc, you have 5 addressing modes for GPRs and FPRS:

    1) base register
    2) base register + index register
    3) base register + offset
    4) auto-update for base register + index register
    5) auto-update for base register + offset register

Now for VSX registers you only have:

    1) base register
    2) base register + index register

So, in the work I'm doing right now, I want to allow reg + offset addressing,
but if the register being loaded is an Altivec register (high part of the VSX
registers), I need to create a secondary reload to load the offset into and
then convert the address to indirect or indexed addressing.  You don't want to
always disallow offset based addressing, but you want to create the secondary
reload when you need to.

Similarly, vector types can do indexed addressing (reg+reg) but if you are
loading or storing the value into GPR registers, you can't do reg+reg
addressing on multi-word items.

Finally, one of the features of ISA 2.07 is the notion of load fusion, where
the hardware will fuse together a load immediate with an adjacent load (for
GPRs, you need the load immediate shifted to be the register that is being
loaded, for VSX registers, you just need the instructions adjacent).  In this
case, before reload we will want to pretend that the machine has addressing to
include the fusion forms, and in secondary reload, you will generate the
combined insn that will become the fusion instruction.

-- 
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460, USA
email: meissner@linux.vnet.ibm.com, phone: +1 (978) 899-4797


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]