https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128545
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Completely untested patch:
--- a/gcc/internal-fn.cc 2026-11-07 19:36:19.412932623 +0100
+++ b/gcc/internal-fn.cc 2026-11-07 20:22:08.441387193 +0100
@@ -6090,6 +6090,56 @@ expand_clmul (machine_mode mode, machine
delete_insns_since (start);
}
}
+ if (CONST_SCALAR_INT_P (op1))
+ {
+ wide_int wop1 = rtx_mode_t (op1, op_mode);
+ unsigned lz = wi::clz (wop1);
+ unsigned tz = wi::ctz (wop1);
+ unsigned cnt = wi::popcount (wop1);
+ if (lz + cnt + tz == op_prec
+ && (pow2p_hwi (cnt) || (lz == 0 && prec == op_prec)))
+ {
+ /* Optimize the cases where one argument is a constant
+ equal to ((((T) 1) << M) - 1) << N for some M and N
+ constants.
+ In this case the clmul (x, ((((T) 1) << M) - 1) << N)
+ result can be computed as
+ T r = x;
+ for (unsigned i = 1; i < M; i = i + i)
+ r ^= x << i;
+ return r << N;
+ If M is not a power of two, it shouldn't be a widening
+ shift and most significant bit of the constant should be
+ set. */
+ rtx x = convert_modes (mode, op_mode, op0, 1);
+ rtx res = gen_reg_rtx (mode);
+ emit_move_insn (res, x);
+ rtx i = gen_reg_rtx (TYPE_MODE (integer_type_node));
+ emit_move_insn (i, const0_rtx);
+ rtx_code_label *loop_label = gen_label_rtx ();
+ emit_label (loop_label);
+ rtx tem = expand_variable_shift (LSHIFT_EXPR, mode, x,
+ make_tree (integer_type_node, i),
+ NULL_RTX, 1);
+ tem = expand_binop (mode, xor_optab, res, tem, res, 1, olw);
+ if (tem != res)
+ emit_move_insn (res, tem);
+ tem = expand_binop (GET_MODE (i), add_optab, i, i, i, 1, olw);
+ if (tem != res)
+ emit_move_insn (i, tem);
+ do_compare_rtx_and_jump (i, gen_int_mode (cnt, GET_MODE (i)),
+ LTU, true, GET_MODE (i), NULL_RTX, NULL,
+ loop_label,
+ profile_probability::very_likely ());
+ if (tz)
+ {
+ tem = expand_shift (LSHIFT_EXPR, mode, res, tz, res, 1);
+ if (tem != res)
+ emit_move_insn (res, tem);
+ }
+ return res;
+ }
+ }
/* Fallback using Kronecker substitution.
Depending on the number of active bits in the operand with smaller
number of those and the result precision we choose smallest spacing
