On Wed, 4 Apr 2018, Jakub Jelinek wrote:
> Hi!
>
> On the following testcase because of the SCCVN limit we end up with a weird,
> but valid, BIT_FIELD_REF - trying to extract V1TImode type out of a V1TImode
> SSA_NAME, with 128 bits width and offset 0 (just SSA_NAME move would be
> enough). Not trying to address why we create it, rather fix how we handle
> it.
>
> The problem is that the BIT_FIELD_REF vector extraction is guarded with:
> (if (VECTOR_TYPE_P (TREE_TYPE (@0))
> && (types_match (type, TREE_TYPE (TREE_TYPE (@0)))
> || (VECTOR_TYPE_P (type)
> && types_match (TREE_TYPE (type), TREE_TYPE (TREE_TYPE (@0))))))
> and thus we must handle vector type as result, but assume incorrectly that
> if count is 1, then we must be asking for an element and thus not vector
> type. For the first hunk we could do what we do in the second hunk, do
> a VIEW_CONVERT_EXPR, but it seems cleaner to just fall through into the
> count > 1 code which builds a CONSTRUCTOR.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
I think I prefer a V_C_E in the first case like in the second. That
may help relaxing the type constraint above to a size one so we
can handle SImode extracts from a VnSFmode vector for example.
It also shows we are lacking a generic
(BIT_FIELD_REF @0 bitsize-of-@0 integer_zerop) to (VIEW_CONVERT @0)
pattern which would need to be ordered before the constructor one
to make sure it matches first (and it would have handled the
very specific case as well).
So - OK with doing the same in the first hunk as in the second.
Thanks,
Richard.
> 2018-04-04 Jakub Jelinek <ja...@redhat.com>
>
> PR middle-end/85195
> * match.pd (BIT_FIELD_REF CONSTRUCTOR@0 @1 @2): If count == 1, only
> use elt value if type is not vector type, otherwise build CONSTRUCTOR.
> For n == const_k, use view_convert.
>
> * gcc.dg/pr85195.c: New test.
>
> --- gcc/match.pd.jj 2018-03-28 21:15:00.000000000 +0200
> +++ gcc/match.pd 2018-04-04 14:35:10.424127155 +0200
> @@ -4648,7 +4648,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (if (multiple_p (idx, k, &elt) && multiple_p (n, k, &count))
> (if (CONSTRUCTOR_NELTS (ctor) == 0)
> { build_constructor (type, NULL); }
> - (if (count == 1)
> + (if (count == 1 && !VECTOR_TYPE_P (type))
> (if (elt < CONSTRUCTOR_NELTS (ctor))
> { CONSTRUCTOR_ELT (ctor, elt)->value; }
> { build_zero_cst (type); })
> @@ -4668,7 +4668,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (if (CONSTRUCTOR_NELTS (ctor) <= idx / const_k)
> { build_zero_cst (type); })
> (if (n == const_k)
> - { CONSTRUCTOR_ELT (ctor, idx / const_k)->value; })
> + (view_convert { CONSTRUCTOR_ELT (ctor, idx / const_k)->value; }))
> (BIT_FIELD_REF { CONSTRUCTOR_ELT (ctor, idx / const_k)->value; }
> @1 { bitsize_int ((idx % const_k) * width); })))))))))
>
> --- gcc/testsuite/gcc.dg/pr85195.c.jj 2018-04-04 14:38:29.233235494 +0200
> +++ gcc/testsuite/gcc.dg/pr85195.c 2018-04-04 14:38:14.696227566 +0200
> @@ -0,0 +1,19 @@
> +/* PR middle-end/85195 */
> +/* { dg-do compile { target int128 } } */
> +/* { dg-options "-Wno-psabi -O -fno-tree-ccp --param=sccvn-max-scc-size=10"
> } */
> +
> +typedef __int128 V __attribute__ ((vector_size (16)));
> +
> +extern int bar (V);
> +
> +V v;
> +int i;
> +
> +V
> +foo (void)
> +{
> + do
> + v *= bar (v & i);
> + while ((V){}[0]);
> + return v;
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
