Re: Support << and >> for offset_int and widest_int

[Richard Sandiford <richard.sandiford@arm.com>]

"H.J. Lu" <hjl.tools@gmail.com> writes:
> On Fri, Apr 29, 2016 at 5:30 AM, Richard Sandiford
> <richard.sandiford@arm.com> wrote:
>> Following on from the comparison patch, I think it makes sense to
>> support << and >> for offset_int (int128_t) and widest_int (intNNN_t),
>> with >> being arithmetic shift.  It doesn't make sense to use
>> logical right shift on a potentially negative offset_int, since
>> the precision of 128 bits has no meaning on the target.
>>
>> Tested on x86_64-linux-gnu and aarch64-linux-gnu.  OK to install?
>>
>> Thanks,
>> Richard
>>
>>
>> gcc/
>>         * wide-int.h: Update offset_int and widest_int documentation.
>>         (WI_SIGNED_SHIFT_RESULT): New macro.
>>         (wi::binary_shift): Define signed_shift_result_type for
>>         shifts on offset_int- and widest_int-like types.
>>         (generic_wide_int): Support <<= and >>= if << and >> are supported.
>>         * tree.h (int_bit_position): Use shift operators instead of wi::
>>          shifts.
>>         * alias.c (adjust_offset_for_component_ref): Likewise.
>>         * expr.c (get_inner_reference): Likewise.
>>         * fold-const.c (fold_comparison): Likewise.
>>         * gimple-fold.c (fold_nonarray_ctor_reference): Likewise.
>>         * gimple-ssa-strength-reduction.c (restructure_reference): Likewise.
>>         * tree-dfa.c (get_ref_base_and_extent): Likewise.
>>         * tree-ssa-alias.c (indirect_ref_may_alias_decl_p): Likewise.
>>         (stmt_kills_ref_p): Likewise.
>>         * tree-ssa-ccp.c (bit_value_binop_1): Likewise.
>>         * tree-ssa-math-opts.c (find_bswap_or_nop_load): Likewise.
>>         * tree-ssa-sccvn.c (copy_reference_ops_from_ref): Likewise.
>>         (ao_ref_init_from_vn_reference): Likewise.
>>
>> gcc/cp/
>>         * init.c (build_new_1): Use shift operators instead of wi:: shifts.
>
> Can you also update change_zero_ext in combine.c:
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70687
>
> It should use wide_int << instead of HOST_WIDE_INT <<< to
> support __int128.

The patch doesn't add wide_int shift operators since wide_ints have no
sign.  You need to use wi:: shift routines for them instead.  However,
there's already a wi::mask function for creating this kind of value
directly.

Like you say, the PR is about converting other code to use wi::, which
is very different from what this patch is doing.  I'll take the PR
anyway though.  Hope to post a patch on Wednesday.

It also looks like the code is missing a check that the ZERO_EXTEND mode
is scalar, since the transformation would be incorrect for vectors.

Thanks,
Richard