[Bug tree-optimization/125405] Missed optimization: mask/sub sign-extension idiom not canonicalized to sign extension

[Milan.Tripkovic@rt-rk.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125405

--- Comment #6 from Milan Tripkovic <Milan.Tripkovic@rt-rk.com> ---
Hello,

Thank you for the suggestions. I updated the pattern incorporating the feedback
from comments #3, #4, and #5:

(simplify
 (minus (bit_and @0 INTEGER_CST@1) (bit_and @0 integer_pow2p@2))
 (if (INTEGRAL_TYPE_P (type)
      && TYPE_PRECISION (type) <= MAX_FIXED_MODE_SIZE
      && wi::to_wide (@1) + 1 == wi::to_wide (@2))
  (with {
    int width = wi::exact_log2 (wi::to_wide (@2)) + 1;
    tree ntype = NULL_TREE;
    if (width > 0 && width < TYPE_PRECISION (type))
      ntype = build_nonstandard_integer_type (width, SIGNED);
   }
   (if (ntype && type_has_mode_precision_p (ntype))
    (convert (convert:ntype @0))))))


I tested the updated pattern on RISC-V (rv64gc)  covering: standard long/int
types, _BitInt with power-of-2 widths (32), non-power-of-2 widths (33, 37, 63),
unsigned BitInt types, negative cases where the pattern should not fire.

Example tests are available at: https://godbolt.org/z/4ndW7rG9W

The asm outputs with my pattern are shown in the comments in the test file.