[PATCH v3 05/12] loongarch: Add an inline roundeven_finite

[Adhemerval Zanella]

Every range reduction that uses roundeven_finite requires a call to
__roundeven.  The FTINTRNE.L.D and FFINT.D.L pair gives the
round-to-nearest-even value in two instructions, independently of the
rounding mode, and exactly for |x| < 2^63.

On a Loongson-3C5000L-LL the CORE-MATH sinf, cosf, and sincosf of the
following patches run 14% to 16% faster with this change (geomean of
the reciprocal-throughput and latency of the benchtests inputs that
preceded the workload rewrite of this series: sinf -14.4%, cosf -15.6%,
sincosf -14.9%).

FTINTRNE raises the inexact exception for a non integer argument,
however the users of roundeven_finite (the range reductions of sinf,
cosf, sincosf, tanf, tanpif, expm1f, sinhf, coshf, gammaf, and the
double erfc, gamma, and lgamma) already produce inexact results for
such arguments anyway.

Checked on loongarch64-linux-gnuf64.
---
 sysdeps/loongarch/fpu/math-roundeven-finite.h | 44 +++++++++++++++++++
 1 file changed, 44 insertions(+)
 create mode 100644 sysdeps/loongarch/fpu/math-roundeven-finite.h

diff --git a/sysdeps/loongarch/fpu/math-roundeven-finite.h b/sysdeps/loongarch/fpu/math-roundeven-finite.h
new file mode 100644
index 00000000000..5d0fdd56b28
--- /dev/null
+++ b/sysdeps/loongarch/fpu/math-roundeven-finite.h
@@ -0,0 +1,44 @@
+/* Round to nearest integer, ties to even, of a finite argument.  LoongArch.
+   Copyright (C) 2026 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#ifndef _MATH_ROUNDEVEN_FINITE_H
+#define _MATH_ROUNDEVEN_FINITE_H 1
+
+/* The result is exact for |x| < 2^63. FTINTRNE also raises the inexact
+   the current usage tolerates it.  */
+static inline double
+roundeven_finite (double x)
+{
+  double r;
+  __asm__ ("ftintrne.l.d %0, %1\n\t"
+	   "ffint.d.l %0, %0"
+	   : "=f" (r) : "f" (x));
+  return r;
+}
+
+static inline float
+roundevenf_finite (float x)
+{
+  float r;
+  __asm__ ("ftintrne.l.s %0, %1\n\t"
+	   "ffint.s.l %0, %0"
+	   : "=f" (r) : "f" (x));
+  return r;
+}
+
+#endif
-- 
2.53.0