https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109393
--- Comment #7 from Manolis Tsamis <manolis.tsamis at vrull dot eu> ---
After some attempts to improve on this, my current solution is:
1) Do not change pointer_int_sum in c-common (otherwise codegen regressions
are observed)
2) Introduce a pattern that folds (unsigned type) (x + CONST_INT_1) *
CONST_INT_2 to (unsigned type) (x * CONST_INT_2) + (unsigned type) (CONST_INT_1
* CONST_INT_1) under the right circumstances.
This combination improves codegen (including the testcase provided) and also
doesn't cause any regressions on the GCC testsuite or benchmarks that I have
tried so far.
Given that a[(ulong) i +- C] in GIMPLE is quite common, I believe this change
helps with folding / canonicalization in many cases.
My current match.pd pattern to do that is below; any feedback or improvements
are welcome.
(simplify
(mult (convert@3 (plus @0 INTEGER_CST@1)) INTEGER_CST@2)
(with { tree tt = TREE_TYPE(@3); }
(if (INTEGRAL_TYPE_P (type)
&& !TYPE_SATURATING (type)
&& !TYPE_OVERFLOW_TRAPS (type)
&& !TYPE_OVERFLOW_SANITIZED (type)
&& !TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
(plus (mult (convert:tt @0) @2) (mult (convert:tt @1) @2)))))
