Re: scalar vector shift expansion problem on 64-bit

[David Miller]

From: David Miller <da...@davemloft.net>
Date: Fri, 28 Oct 2011 01:05:54 -0400 (EDT)

> So should expand_vector_broadcast() really provide this invariant to
> the vec_init expander, or does the vec_init expander need to tidy
> things up with gen_lowpart() etc. calls?

Richard I don't know if you had a chance to look into this at all yet,
but I wanted to make a comment about vec_init in general.

I've come to find that I want the compiler to do as little as possible
with the expressions that get put into vector initializers.

I don't want it to modify the mode of the individual inner elements in
the assignment. I also don't want it to force mems into registers.

In fact, the less it does the better.

I want to make use of the special VIS load instructions that can take
a QImode or HImode value in memory and load it zero extended into a
64-bit float register.

For example:

int x;

__v8qi test_v8qi(void)
{
  __v8qi ret = { x, x, x, x, x, x, x, x };

  return ret;
}

I want to be able to generate:

test_v8qi:
        sethi     %hi(x + 3), %g1
        or        %g1, %lo(x + 3), %g1
        ldda      [%g1] ASI_FL8_P, %f2
        sethi     %hi(0x77777777), %g2
        or        %g2, %lo(0x77777777), %g2
        bmask     %g2, %g0, %g0
        retl
         bshuffle  %f2, %f2, %f0

but I can't because the vec_init expander sees:

(parallel:V8QI [
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
        (reg:QI 110 [ D.2249 ])
    ])

in operands[1].