[Bug tree-optimization/45397] [5/6/7 Regression] Issues with integer narrowing conversions

[law at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=45397

Jeffrey A. Law <law at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |law at redhat dot com

--- Comment #21 from Jeffrey A. Law <law at redhat dot com> ---
Couldn't we solve this with a pattern like this in match.pd:

(simplify
  (convert (plus:c @0 (convert @1)))
  (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
       && INTEGRAL_TYPE_P (TREE_TYPE (@0))
       && INTEGRAL_TYPE_P (type)
       && types_match (@1, type)
       && (TYPE_PRECISION (TREE_TYPE (@1)) > TYPE_PRECISION (TREE_TYPE (@0))))
    (bit_and (plus (convert @0) @1)
       { wide_int_to_tree
                (type, wi::mask (TYPE_PRECISION (TREE_TYPE (@0)), false,
TYPE_PRECISION (type))); })))

Essentially this widens the arithmetic, then masks off undesirable bits in the
result.  As written I think this doesn't reduce the number of expressions
consistently even though it helps the testcase.

The reason it helps the testcase is the (convert @0) in the output pattern
turns out to be a common subexpression.  If (convert @0) is CSE-able, then this
does reduce the number of expressions.  Though we probably want to avoid
widening beyond a target's wordsize (which I loathe to test for).

Thoughts anyone?