Re: [PATCH v2] rs6000: Add new pass for replacement of contiguous addresses vector load lxv with lxvp

[Michael Meissner <meissner@linux.ibm.com>]

On Tue, Nov 28, 2023 at 05:44:43PM +0800, Kewen.Lin wrote:
> on 2023/11/28 15:05, Michael Meissner wrote:
> > I tried using this patch to compare with the vector size attribute patch I
> > posted.  I could not build it as a cross compiler on my x86_64 because the
> > assembler gives the following error:
> > 
> > Error: operand out of domain (11 is not a multiple of 2) for
> > std_stacktrace-elf.o.  If you look at the assembler, it has combined a lxvp 11
> > and lxvp 12 into:
> > 
> >         lxvp 11,0(9)
> > 
> > The powerpc architecture requires that registers that are loaded with load
> > vector pair and stored with store vector point instructions only load/store
> > even/odd register pairs, and not odd/even pairs.  Unfortunately, it will mean
> > that this optimization will match less often.
> > 
> 
> Yes, the current implementation need some refinements, as comments in [1]:
> 
> > Besides, it seems a bad idea to put this pass after reload? as register allocation
> > finishes, this pairing has to be restricted by the reg No. (I didn't see any
> > checking on the reg No. relationship for paring btw.)
> > 
> > Looking forward to the comments from Segher/David/Peter/Mike etc.
> 
> I wonder if we should consider running such pass before reload instead.
> 
> [1] https://gcc.gnu.org/pipermail/gcc-patches/2023-November/638070.html
> 
> BR,
> Kewen

If I add code to check if the target register is even, then the following
fails:

/home/meissner/fsf-src/work148-ajit/libquadmath/math/erfq.c: In function ‘erfcq’:
/home/meissner/fsf-src/work148-ajit/libquadmath/math/erfq.c:943:1: error: insn does not satisfy its constraints:
  943 | }
      | ^
(insn 1087 1939 1088 66 (set (reg/v:KF 74 10 [orig:643 y ] [643])
        (fma:KF (reg/v:KF 64 0 [orig:153 z ] [153])
            (reg/v:KF 65 1 [orig:639 y ] [639])
            (reg:KF 76 12 [orig:642 MEM[(const _Float128 *)p_276 + 16B] ] [642]))) "/home/meissner/fsf-src/work148-ajit/libquadmath/math/erfq.c":112:9 1004 {fmakf4_hw}
     (expr_list:REG_DEAD (reg/v:KF 65 1 [orig:639 y ] [639])
        (nil)))

In particular, the IEEE 128-bit arithmetic functions require Altivec registers.
So we would need to make sure the new insns all meet their constraints.

I tend to think that it would be desirable to do it before reload.  But then we
will need to check if extra moves are generated.  I suspect we will need
Peter's patch to allow 128-bit types that are subregs of OOmode.  I.e., the
code generated would change:

	(set (reg:MODE1 tmp-reg)
	     (mem ...+8))

	(set (reg:MODE2 tmp-reg+1)
	     (mem ...))

to:

	(set (reg:OO vp-reg)
	     (mem ...))

	(set (reg:MODE1 tmp-reg)
	     (subreg:MODE1 (reg:OO vp-reg 0)))

	(set (reg:MODE2 tmp-reg+1)
	     (subreg:MODE2 (reg:OO vp-reg 16)))

Note, I may have the offsets and register numbers backwards in terms of endian.

-- 
Michael Meissner, IBM
PO Box 98, Ayer, Massachusetts, USA, 01432
email: meissner@linux.ibm.com