aarch64: Add vector implementations of log1p routines

May discard sign of zero.

13 files changed, 500 insertions, 0 deletions

diff --git a/sysdeps/aarch64/fpu/Makefile b/sysdeps/aarch64/fpu/Makefile
index 364efbeac1..c77c709edd 100644
--- a/sysdeps/aarch64/fpu/Makefile
+++ b/sysdeps/aarch64/fpu/Makefile
@@ -8,6 +8,7 @@ libmvec-supported-funcs = acos \
                           exp2 \
                           log \
                           log10 \
+                          log1p \
                           log2 \
                           sin \
                           tan
diff --git a/sysdeps/aarch64/fpu/Versions b/sysdeps/aarch64/fpu/Versions
index 99492b3d33..2543649fbe 100644
--- a/sysdeps/aarch64/fpu/Versions
+++ b/sysdeps/aarch64/fpu/Versions
@@ -46,6 +46,10 @@ libmvec {
     _ZGVnN2v_log10;
     _ZGVsMxv_log10f;
     _ZGVsMxv_log10;
+    _ZGVnN4v_log1pf;
+    _ZGVnN2v_log1p;
+    _ZGVsMxv_log1pf;
+    _ZGVsMxv_log1p;
     _ZGVnN4v_log2f;
     _ZGVnN2v_log2;
     _ZGVsMxv_log2f;
diff --git a/sysdeps/aarch64/fpu/bits/math-vector.h b/sysdeps/aarch64/fpu/bits/math-vector.h
index 7666c09083..51915cef22 100644
--- a/sysdeps/aarch64/fpu/bits/math-vector.h
+++ b/sysdeps/aarch64/fpu/bits/math-vector.h
@@ -59,6 +59,7 @@ __vpcs __f32x4_t _ZGVnN4v_exp10f (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_exp2f (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_logf (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_log10f (__f32x4_t);
+__vpcs __f32x4_t _ZGVnN4v_log1pf (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_log2f (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_sinf (__f32x4_t);
 __vpcs __f32x4_t _ZGVnN4v_tanf (__f32x4_t);
@@ -73,6 +74,7 @@ __vpcs __f64x2_t _ZGVnN2v_exp10 (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_exp2 (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_log (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_log10 (__f64x2_t);
+__vpcs __f64x2_t _ZGVnN2v_log1p (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_log2 (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_sin (__f64x2_t);
 __vpcs __f64x2_t _ZGVnN2v_tan (__f64x2_t);
@@ -92,6 +94,7 @@ __sv_f32_t _ZGVsMxv_exp10f (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_exp2f (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_logf (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_log10f (__sv_f32_t, __sv_bool_t);
+__sv_f32_t _ZGVsMxv_log1pf (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_log2f (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_sinf (__sv_f32_t, __sv_bool_t);
 __sv_f32_t _ZGVsMxv_tanf (__sv_f32_t, __sv_bool_t);
@@ -106,6 +109,7 @@ __sv_f64_t _ZGVsMxv_exp10 (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_exp2 (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_log (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_log10 (__sv_f64_t, __sv_bool_t);
+__sv_f64_t _ZGVsMxv_log1p (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_log2 (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_sin (__sv_f64_t, __sv_bool_t);
 __sv_f64_t _ZGVsMxv_tan (__sv_f64_t, __sv_bool_t);
diff --git a/sysdeps/aarch64/fpu/log1p_advsimd.c b/sysdeps/aarch64/fpu/log1p_advsimd.c
new file mode 100644
index 0000000000..a117e1b6dc
--- /dev/null
+++ b/sysdeps/aarch64/fpu/log1p_advsimd.c
@@ -0,0 +1,129 @@
+/* Double-precision AdvSIMD log1p
+
+   Copyright (C) 2023 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/>.  */
+
+#include "v_math.h"
+#include "poly_advsimd_f64.h"
+
+const static struct data
+{
+  float64x2_t poly[19], ln2[2];
+  uint64x2_t hf_rt2_top, one_m_hf_rt2_top, umask, inf, minus_one;
+  int64x2_t one_top;
+} data = {
+  /* Generated using Remez, deg=20, in [sqrt(2)/2-1, sqrt(2)-1].  */
+  .poly = { V2 (-0x1.ffffffffffffbp-2), V2 (0x1.55555555551a9p-2),
+	    V2 (-0x1.00000000008e3p-2), V2 (0x1.9999999a32797p-3),
+	    V2 (-0x1.555555552fecfp-3), V2 (0x1.249248e071e5ap-3),
+	    V2 (-0x1.ffffff8bf8482p-4), V2 (0x1.c71c8f07da57ap-4),
+	    V2 (-0x1.9999ca4ccb617p-4), V2 (0x1.7459ad2e1dfa3p-4),
+	    V2 (-0x1.554d2680a3ff2p-4), V2 (0x1.3b4c54d487455p-4),
+	    V2 (-0x1.2548a9ffe80e6p-4), V2 (0x1.0f389a24b2e07p-4),
+	    V2 (-0x1.eee4db15db335p-5), V2 (0x1.e95b494d4a5ddp-5),
+	    V2 (-0x1.15fdf07cb7c73p-4), V2 (0x1.0310b70800fcfp-4),
+	    V2 (-0x1.cfa7385bdb37ep-6) },
+  .ln2 = { V2 (0x1.62e42fefa3800p-1), V2 (0x1.ef35793c76730p-45) },
+  /* top32(asuint64(sqrt(2)/2)) << 32.  */
+  .hf_rt2_top = V2 (0x3fe6a09e00000000),
+  /* (top32(asuint64(1)) - top32(asuint64(sqrt(2)/2))) << 32.  */
+  .one_m_hf_rt2_top = V2 (0x00095f6200000000),
+  .umask = V2 (0x000fffff00000000),
+  .one_top = V2 (0x3ff),
+  .inf = V2 (0x7ff0000000000000),
+  .minus_one = V2 (0xbff0000000000000)
+};
+
+#define BottomMask v_u64 (0xffffffff)
+
+static float64x2_t VPCS_ATTR NOINLINE
+special_case (float64x2_t x, float64x2_t y, uint64x2_t special)
+{
+  return v_call_f64 (log1p, x, y, special);
+}
+
+/* Vector log1p approximation using polynomial on reduced interval. Routine is
+   a modification of the algorithm used in scalar log1p, with no shortcut for
+   k=0 and no narrowing for f and k. Maximum observed error is 2.45 ULP:
+   _ZGVnN2v_log1p(0x1.658f7035c4014p+11) got 0x1.fd61d0727429dp+2
+					want 0x1.fd61d0727429fp+2 .  */
+VPCS_ATTR float64x2_t V_NAME_D1 (log1p) (float64x2_t x)
+{
+  const struct data *d = ptr_barrier (&data);
+  uint64x2_t ix = vreinterpretq_u64_f64 (x);
+  uint64x2_t ia = vreinterpretq_u64_f64 (vabsq_f64 (x));
+  uint64x2_t special = vcgeq_u64 (ia, d->inf);
+
+#if WANT_SIMD_EXCEPT
+  special = vorrq_u64 (special,
+		       vcgeq_u64 (ix, vreinterpretq_u64_f64 (v_f64 (-1))));
+  if (__glibc_unlikely (v_any_u64 (special)))
+    x = v_zerofy_f64 (x, special);
+#else
+  special = vorrq_u64 (special, vcleq_f64 (x, v_f64 (-1)));
+#endif
+
+  /* With x + 1 = t * 2^k (where t = f + 1 and k is chosen such that f
+			   is in [sqrt(2)/2, sqrt(2)]):
+     log1p(x) = k*log(2) + log1p(f).
+
+     f may not be representable exactly, so we need a correction term:
+     let m = round(1 + x), c = (1 + x) - m.
+     c << m: at very small x, log1p(x) ~ x, hence:
+     log(1+x) - log(m) ~ c/m.
+
+     We therefore calculate log1p(x) by k*log2 + log1p(f) + c/m.  */
+
+  /* Obtain correctly scaled k by manipulation in the exponent.
+     The scalar algorithm casts down to 32-bit at this point to calculate k and
+     u_red. We stay in double-width to obtain f and k, using the same constants
+     as the scalar algorithm but shifted left by 32.  */
+  float64x2_t m = vaddq_f64 (x, v_f64 (1));
+  uint64x2_t mi = vreinterpretq_u64_f64 (m);
+  uint64x2_t u = vaddq_u64 (mi, d->one_m_hf_rt2_top);
+
+  int64x2_t ki
+      = vsubq_s64 (vreinterpretq_s64_u64 (vshrq_n_u64 (u, 52)), d->one_top);
+  float64x2_t k = vcvtq_f64_s64 (ki);
+
+  /* Reduce x to f in [sqrt(2)/2, sqrt(2)].  */
+  uint64x2_t utop = vaddq_u64 (vandq_u64 (u, d->umask), d->hf_rt2_top);
+  uint64x2_t u_red = vorrq_u64 (utop, vandq_u64 (mi, BottomMask));
+  float64x2_t f = vsubq_f64 (vreinterpretq_f64_u64 (u_red), v_f64 (1));
+
+  /* Correction term c/m.  */
+  float64x2_t cm = vdivq_f64 (vsubq_f64 (x, vsubq_f64 (m, v_f64 (1))), m);
+
+  /* Approximate log1p(x) on the reduced input using a polynomial. Because
+     log1p(0)=0 we choose an approximation of the form:
+       x + C0*x^2 + C1*x^3 + C2x^4 + ...
+     Hence approximation has the form f + f^2 * P(f)
+      where P(x) = C0 + C1*x + C2x^2 + ...
+     Assembling this all correctly is dealt with at the final step.  */
+  float64x2_t f2 = vmulq_f64 (f, f);
+  float64x2_t p = v_pw_horner_18_f64 (f, f2, d->poly);
+
+  float64x2_t ylo = vfmaq_f64 (cm, k, d->ln2[1]);
+  float64x2_t yhi = vfmaq_f64 (f, k, d->ln2[0]);
+  float64x2_t y = vaddq_f64 (ylo, yhi);
+
+  if (__glibc_unlikely (v_any_u64 (special)))
+    return special_case (vreinterpretq_f64_u64 (ix), vfmaq_f64 (y, f2, p),
+			 special);
+
+  return vfmaq_f64 (y, f2, p);
+}
diff --git a/sysdeps/aarch64/fpu/log1p_sve.c b/sysdeps/aarch64/fpu/log1p_sve.c
new file mode 100644
index 0000000000..169156748d
--- /dev/null
+++ b/sysdeps/aarch64/fpu/log1p_sve.c
@@ -0,0 +1,118 @@
+/* Double-precision SVE log1p
+
+   Copyright (C) 2023 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/>.  */
+
+#include "sv_math.h"
+#include "poly_sve_f64.h"
+
+static const struct data
+{
+  double poly[19];
+  double ln2_hi, ln2_lo;
+  uint64_t hfrt2_top, onemhfrt2_top, inf, mone;
+} data = {
+  /* Generated using Remez in [ sqrt(2)/2 - 1, sqrt(2) - 1]. Order 20
+     polynomial, however first 2 coefficients are 0 and 1 so are not stored.  */
+  .poly = { -0x1.ffffffffffffbp-2, 0x1.55555555551a9p-2, -0x1.00000000008e3p-2,
+	    0x1.9999999a32797p-3, -0x1.555555552fecfp-3, 0x1.249248e071e5ap-3,
+	    -0x1.ffffff8bf8482p-4, 0x1.c71c8f07da57ap-4, -0x1.9999ca4ccb617p-4,
+	    0x1.7459ad2e1dfa3p-4, -0x1.554d2680a3ff2p-4, 0x1.3b4c54d487455p-4,
+	    -0x1.2548a9ffe80e6p-4, 0x1.0f389a24b2e07p-4, -0x1.eee4db15db335p-5,
+	    0x1.e95b494d4a5ddp-5, -0x1.15fdf07cb7c73p-4, 0x1.0310b70800fcfp-4,
+	    -0x1.cfa7385bdb37ep-6, },
+  .ln2_hi = 0x1.62e42fefa3800p-1,
+  .ln2_lo = 0x1.ef35793c76730p-45,
+  /* top32(asuint64(sqrt(2)/2)) << 32.  */
+  .hfrt2_top = 0x3fe6a09e00000000,
+  /* (top32(asuint64(1)) - top32(asuint64(sqrt(2)/2))) << 32.  */
+  .onemhfrt2_top = 0x00095f6200000000,
+  .inf = 0x7ff0000000000000,
+  .mone = 0xbff0000000000000,
+};
+
+#define AbsMask 0x7fffffffffffffff
+#define BottomMask 0xffffffff
+
+static svfloat64_t NOINLINE
+special_case (svbool_t special, svfloat64_t x, svfloat64_t y)
+{
+  return sv_call_f64 (log1p, x, y, special);
+}
+
+/* Vector approximation for log1p using polynomial on reduced interval. Maximum
+   observed error is 2.46 ULP:
+   _ZGVsMxv_log1p(0x1.654a1307242a4p+11) got 0x1.fd5565fb590f4p+2
+					want 0x1.fd5565fb590f6p+2.  */
+svfloat64_t SV_NAME_D1 (log1p) (svfloat64_t x, svbool_t pg)
+{
+  const struct data *d = ptr_barrier (&data);
+  svuint64_t ix = svreinterpret_u64 (x);
+  svuint64_t ax = svand_x (pg, ix, AbsMask);
+  svbool_t special
+      = svorr_z (pg, svcmpge (pg, ax, d->inf), svcmpge (pg, ix, d->mone));
+
+  /* With x + 1 = t * 2^k (where t = f + 1 and k is chosen such that f
+			   is in [sqrt(2)/2, sqrt(2)]):
+     log1p(x) = k*log(2) + log1p(f).
+
+     f may not be representable exactly, so we need a correction term:
+     let m = round(1 + x), c = (1 + x) - m.
+     c << m: at very small x, log1p(x) ~ x, hence:
+     log(1+x) - log(m) ~ c/m.
+
+     We therefore calculate log1p(x) by k*log2 + log1p(f) + c/m.  */
+
+  /* Obtain correctly scaled k by manipulation in the exponent.
+     The scalar algorithm casts down to 32-bit at this point to calculate k and
+     u_red. We stay in double-width to obtain f and k, using the same constants
+     as the scalar algorithm but shifted left by 32.  */
+  svfloat64_t m = svadd_x (pg, x, 1);
+  svuint64_t mi = svreinterpret_u64 (m);
+  svuint64_t u = svadd_x (pg, mi, d->onemhfrt2_top);
+
+  svint64_t ki = svsub_x (pg, svreinterpret_s64 (svlsr_x (pg, u, 52)), 0x3ff);
+  svfloat64_t k = svcvt_f64_x (pg, ki);
+
+  /* Reduce x to f in [sqrt(2)/2, sqrt(2)].  */
+  svuint64_t utop
+      = svadd_x (pg, svand_x (pg, u, 0x000fffff00000000), d->hfrt2_top);
+  svuint64_t u_red = svorr_x (pg, utop, svand_x (pg, mi, BottomMask));
+  svfloat64_t f = svsub_x (pg, svreinterpret_f64 (u_red), 1);
+
+  /* Correction term c/m.  */
+  svfloat64_t cm = svdiv_x (pg, svsub_x (pg, x, svsub_x (pg, m, 1)), m);
+
+  /* Approximate log1p(x) on the reduced input using a polynomial. Because
+     log1p(0)=0 we choose an approximation of the form:
+	x + C0*x^2 + C1*x^3 + C2x^4 + ...
+     Hence approximation has the form f + f^2 * P(f)
+     where P(x) = C0 + C1*x + C2x^2 + ...
+     Assembling this all correctly is dealt with at the final step.  */
+  svfloat64_t f2 = svmul_x (pg, f, f), f4 = svmul_x (pg, f2, f2),
+	      f8 = svmul_x (pg, f4, f4), f16 = svmul_x (pg, f8, f8);
+  svfloat64_t p = sv_estrin_18_f64_x (pg, f, f2, f4, f8, f16, d->poly);
+
+  svfloat64_t ylo = svmla_x (pg, cm, k, d->ln2_lo);
+  svfloat64_t yhi = svmla_x (pg, f, k, d->ln2_hi);
+  svfloat64_t y = svmla_x (pg, svadd_x (pg, ylo, yhi), f2, p);
+
+  if (__glibc_unlikely (svptest_any (pg, special)))
+    return special_case (special, x, y);
+
+  return y;
+}
diff --git a/sysdeps/aarch64/fpu/log1pf_advsimd.c b/sysdeps/aarch64/fpu/log1pf_advsimd.c
new file mode 100644
index 0000000000..3748830de8
--- /dev/null
+++ b/sysdeps/aarch64/fpu/log1pf_advsimd.c
@@ -0,0 +1,128 @@
+/* Single-precision AdvSIMD log1p
+
+   Copyright (C) 2023 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/>.  */
+
+#include "v_math.h"
+#include "poly_advsimd_f32.h"
+
+const static struct data
+{
+  float32x4_t poly[8], ln2;
+  uint32x4_t tiny_bound, minus_one, four, thresh;
+  int32x4_t three_quarters;
+} data = {
+  .poly = { /* Generated using FPMinimax in [-0.25, 0.5]. First two coefficients
+	       (1, -0.5) are not stored as they can be generated more
+	       efficiently.  */
+	    V4 (0x1.5555aap-2f), V4 (-0x1.000038p-2f), V4 (0x1.99675cp-3f),
+	    V4 (-0x1.54ef78p-3f), V4 (0x1.28a1f4p-3f), V4 (-0x1.0da91p-3f),
+	    V4 (0x1.abcb6p-4f), V4 (-0x1.6f0d5ep-5f) },
+  .ln2 = V4 (0x1.62e43p-1f),
+  .tiny_bound = V4 (0x34000000), /* asuint32(0x1p-23). ulp=0.5 at 0x1p-23.  */
+  .thresh = V4 (0x4b800000), /* asuint32(INFINITY) - tiny_bound.  */
+  .minus_one = V4 (0xbf800000),
+  .four = V4 (0x40800000),
+  .three_quarters = V4 (0x3f400000)
+};
+
+static inline float32x4_t
+eval_poly (float32x4_t m, const float32x4_t *p)
+{
+  /* Approximate log(1+m) on [-0.25, 0.5] using split Estrin scheme.  */
+  float32x4_t p_12 = vfmaq_f32 (v_f32 (-0.5), m, p[0]);
+  float32x4_t p_34 = vfmaq_f32 (p[1], m, p[2]);
+  float32x4_t p_56 = vfmaq_f32 (p[3], m, p[4]);
+  float32x4_t p_78 = vfmaq_f32 (p[5], m, p[6]);
+
+  float32x4_t m2 = vmulq_f32 (m, m);
+  float32x4_t p_02 = vfmaq_f32 (m, m2, p_12);
+  float32x4_t p_36 = vfmaq_f32 (p_34, m2, p_56);
+  float32x4_t p_79 = vfmaq_f32 (p_78, m2, p[7]);
+
+  float32x4_t m4 = vmulq_f32 (m2, m2);
+  float32x4_t p_06 = vfmaq_f32 (p_02, m4, p_36);
+  return vfmaq_f32 (p_06, m4, vmulq_f32 (m4, p_79));
+}
+
+static float32x4_t NOINLINE VPCS_ATTR
+special_case (float32x4_t x, float32x4_t y, uint32x4_t special)
+{
+  return v_call_f32 (log1pf, x, y, special);
+}
+
+/* Vector log1pf approximation using polynomial on reduced interval. Accuracy
+   is roughly 2.02 ULP:
+   log1pf(0x1.21e13ap-2) got 0x1.fe8028p-3 want 0x1.fe802cp-3.  */
+VPCS_ATTR float32x4_t V_NAME_F1 (log1p) (float32x4_t x)
+{
+  const struct data *d = ptr_barrier (&data);
+
+  uint32x4_t ix = vreinterpretq_u32_f32 (x);
+  uint32x4_t ia = vreinterpretq_u32_f32 (vabsq_f32 (x));
+  uint32x4_t special_cases
+      = vorrq_u32 (vcgeq_u32 (vsubq_u32 (ia, d->tiny_bound), d->thresh),
+		   vcgeq_u32 (ix, d->minus_one));
+  float32x4_t special_arg = x;
+
+#if WANT_SIMD_EXCEPT
+  if (__glibc_unlikely (v_any_u32 (special_cases)))
+    /* Side-step special lanes so fenv exceptions are not triggered
+       inadvertently.  */
+    x = v_zerofy_f32 (x, special_cases);
+#endif
+
+  /* With x + 1 = t * 2^k (where t = m + 1 and k is chosen such that m
+			   is in [-0.25, 0.5]):
+     log1p(x) = log(t) + log(2^k) = log1p(m) + k*log(2).
+
+     We approximate log1p(m) with a polynomial, then scale by
+     k*log(2). Instead of doing this directly, we use an intermediate
+     scale factor s = 4*k*log(2) to ensure the scale is representable
+     as a normalised fp32 number.  */
+
+  float32x4_t m = vaddq_f32 (x, v_f32 (1.0f));
+
+  /* Choose k to scale x to the range [-1/4, 1/2].  */
+  int32x4_t k
+      = vandq_s32 (vsubq_s32 (vreinterpretq_s32_f32 (m), d->three_quarters),
+		   v_s32 (0xff800000));
+  uint32x4_t ku = vreinterpretq_u32_s32 (k);
+
+  /* Scale x by exponent manipulation.  */
+  float32x4_t m_scale
+      = vreinterpretq_f32_u32 (vsubq_u32 (vreinterpretq_u32_f32 (x), ku));
+
+  /* Scale up to ensure that the scale factor is representable as normalised
+     fp32 number, and scale m down accordingly.  */
+  float32x4_t s = vreinterpretq_f32_u32 (vsubq_u32 (d->four, ku));
+  m_scale = vaddq_f32 (m_scale, vfmaq_f32 (v_f32 (-1.0f), v_f32 (0.25f), s));
+
+  /* Evaluate polynomial on the reduced interval.  */
+  float32x4_t p = eval_poly (m_scale, d->poly);
+
+  /* The scale factor to be applied back at the end - by multiplying float(k)
+     by 2^-23 we get the unbiased exponent of k.  */
+  float32x4_t scale_back = vcvtq_f32_s32 (vshrq_n_s32 (k, 23));
+
+  /* Apply the scaling back.  */
+  float32x4_t y = vfmaq_f32 (p, scale_back, d->ln2);
+
+  if (__glibc_unlikely (v_any_u32 (special_cases)))
+    return special_case (special_arg, y, special_cases);
+  return y;
+}
diff --git a/sysdeps/aarch64/fpu/log1pf_sve.c b/sysdeps/aarch64/fpu/log1pf_sve.c
new file mode 100644
index 0000000000..712f62b9ce
--- /dev/null
+++ b/sysdeps/aarch64/fpu/log1pf_sve.c
@@ -0,0 +1,100 @@
+/* Single-precision SVE log1p
+
+   Copyright (C) 2023 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/>.  */
+
+#include "sv_math.h"
+#include "poly_sve_f32.h"
+
+static const struct data
+{
+  float poly[8];
+  float ln2, exp_bias;
+  uint32_t four, three_quarters;
+} data = {.poly = {/* Do not store first term of polynomial, which is -0.5, as
+                      this can be fmov-ed directly instead of including it in
+                      the main load-and-mla polynomial schedule.  */
+		   0x1.5555aap-2f, -0x1.000038p-2f, 0x1.99675cp-3f,
+		   -0x1.54ef78p-3f, 0x1.28a1f4p-3f, -0x1.0da91p-3f,
+		   0x1.abcb6p-4f, -0x1.6f0d5ep-5f},
+	  .ln2 = 0x1.62e43p-1f,
+	  .exp_bias = 0x1p-23f,
+	  .four = 0x40800000,
+	  .three_quarters = 0x3f400000};
+
+#define SignExponentMask 0xff800000
+
+static svfloat32_t NOINLINE
+special_case (svfloat32_t x, svfloat32_t y, svbool_t special)
+{
+  return sv_call_f32 (log1pf, x, y, special);
+}
+
+/* Vector log1pf approximation using polynomial on reduced interval. Worst-case
+   error is 1.27 ULP very close to 0.5.
+   _ZGVsMxv_log1pf(0x1.fffffep-2) got 0x1.9f324p-2
+				 want 0x1.9f323ep-2.  */
+svfloat32_t SV_NAME_F1 (log1p) (svfloat32_t x, svbool_t pg)
+{
+  const struct data *d = ptr_barrier (&data);
+  /* x < -1, Inf/Nan.  */
+  svbool_t special = svcmpeq (pg, svreinterpret_u32 (x), 0x7f800000);
+  special = svorn_z (pg, special, svcmpge (pg, x, -1));
+
+  /* With x + 1 = t * 2^k (where t = m + 1 and k is chosen such that m
+			   is in [-0.25, 0.5]):
+     log1p(x) = log(t) + log(2^k) = log1p(m) + k*log(2).
+
+     We approximate log1p(m) with a polynomial, then scale by
+     k*log(2). Instead of doing this directly, we use an intermediate
+     scale factor s = 4*k*log(2) to ensure the scale is representable
+     as a normalised fp32 number.  */
+  svfloat32_t m = svadd_x (pg, x, 1);
+
+  /* Choose k to scale x to the range [-1/4, 1/2].  */
+  svint32_t k
+      = svand_x (pg, svsub_x (pg, svreinterpret_s32 (m), d->three_quarters),
+		 sv_s32 (SignExponentMask));
+
+  /* Scale x by exponent manipulation.  */
+  svfloat32_t m_scale = svreinterpret_f32 (
+      svsub_x (pg, svreinterpret_u32 (x), svreinterpret_u32 (k)));
+
+  /* Scale up to ensure that the scale factor is representable as normalised
+     fp32 number, and scale m down accordingly.  */
+  svfloat32_t s = svreinterpret_f32 (svsubr_x (pg, k, d->four));
+  m_scale = svadd_x (pg, m_scale, svmla_x (pg, sv_f32 (-1), s, 0.25));
+
+  /* Evaluate polynomial on reduced interval.  */
+  svfloat32_t ms2 = svmul_x (pg, m_scale, m_scale),
+	      ms4 = svmul_x (pg, ms2, ms2);
+  svfloat32_t p = sv_estrin_7_f32_x (pg, m_scale, ms2, ms4, d->poly);
+  p = svmad_x (pg, m_scale, p, -0.5);
+  p = svmla_x (pg, m_scale, m_scale, svmul_x (pg, m_scale, p));
+
+  /* The scale factor to be applied back at the end - by multiplying float(k)
+     by 2^-23 we get the unbiased exponent of k.  */
+  svfloat32_t scale_back = svmul_x (pg, svcvt_f32_x (pg, k), d->exp_bias);
+
+  /* Apply the scaling back.  */
+  svfloat32_t y = svmla_x (pg, p, scale_back, d->ln2);
+
+  if (__glibc_unlikely (svptest_any (pg, special)))
+    return special_case (x, y, special);
+
+  return y;
+}
diff --git a/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c b/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
index 0ac0240171..fc9e7aec47 100644
--- a/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-advsimd-wrappers.c
@@ -33,6 +33,7 @@ VPCS_VECTOR_WRAPPER (exp10_advsimd, _ZGVnN2v_exp10)
 VPCS_VECTOR_WRAPPER (exp2_advsimd, _ZGVnN2v_exp2)
 VPCS_VECTOR_WRAPPER (log_advsimd, _ZGVnN2v_log)
 VPCS_VECTOR_WRAPPER (log10_advsimd, _ZGVnN2v_log10)
+VPCS_VECTOR_WRAPPER (log1p_advsimd, _ZGVnN2v_log1p)
 VPCS_VECTOR_WRAPPER (log2_advsimd, _ZGVnN2v_log2)
 VPCS_VECTOR_WRAPPER (sin_advsimd, _ZGVnN2v_sin)
 VPCS_VECTOR_WRAPPER (tan_advsimd, _ZGVnN2v_tan)
diff --git a/sysdeps/aarch64/fpu/test-double-sve-wrappers.c b/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
index 5bbc4d58c1..aea589d5fb 100644
--- a/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-double-sve-wrappers.c
@@ -52,6 +52,7 @@ SVE_VECTOR_WRAPPER (exp10_sve, _ZGVsMxv_exp10)
 SVE_VECTOR_WRAPPER (exp2_sve, _ZGVsMxv_exp2)
 SVE_VECTOR_WRAPPER (log_sve, _ZGVsMxv_log)
 SVE_VECTOR_WRAPPER (log10_sve, _ZGVsMxv_log10)
+SVE_VECTOR_WRAPPER (log1p_sve, _ZGVsMxv_log1p)
 SVE_VECTOR_WRAPPER (log2_sve, _ZGVsMxv_log2)
 SVE_VECTOR_WRAPPER (sin_sve, _ZGVsMxv_sin)
 SVE_VECTOR_WRAPPER (tan_sve, _ZGVsMxv_tan)
diff --git a/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c b/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
index a557bfc3a6..446fd7f538 100644
--- a/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-advsimd-wrappers.c
@@ -33,6 +33,7 @@ VPCS_VECTOR_WRAPPER (exp10f_advsimd, _ZGVnN4v_exp10f)
 VPCS_VECTOR_WRAPPER (exp2f_advsimd, _ZGVnN4v_exp2f)
 VPCS_VECTOR_WRAPPER (logf_advsimd, _ZGVnN4v_logf)
 VPCS_VECTOR_WRAPPER (log10f_advsimd, _ZGVnN4v_log10f)
+VPCS_VECTOR_WRAPPER (log1pf_advsimd, _ZGVnN4v_log1pf)
 VPCS_VECTOR_WRAPPER (log2f_advsimd, _ZGVnN4v_log2f)
 VPCS_VECTOR_WRAPPER (sinf_advsimd, _ZGVnN4v_sinf)
 VPCS_VECTOR_WRAPPER (tanf_advsimd, _ZGVnN4v_tanf)
diff --git a/sysdeps/aarch64/fpu/test-float-sve-wrappers.c b/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
index f36939e2c4..ac17f60856 100644
--- a/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
+++ b/sysdeps/aarch64/fpu/test-float-sve-wrappers.c
@@ -52,6 +52,7 @@ SVE_VECTOR_WRAPPER (exp10f_sve, _ZGVsMxv_exp10f)
 SVE_VECTOR_WRAPPER (exp2f_sve, _ZGVsMxv_exp2f)
 SVE_VECTOR_WRAPPER (logf_sve, _ZGVsMxv_logf)
 SVE_VECTOR_WRAPPER (log10f_sve, _ZGVsMxv_log10f)
+SVE_VECTOR_WRAPPER (log1pf_sve, _ZGVsMxv_log1pf)
 SVE_VECTOR_WRAPPER (log2f_sve, _ZGVsMxv_log2f)
 SVE_VECTOR_WRAPPER (sinf_sve, _ZGVsMxv_sinf)
 SVE_VECTOR_WRAPPER (tanf_sve, _ZGVsMxv_tanf)
diff --git a/sysdeps/aarch64/libm-test-ulps b/sysdeps/aarch64/libm-test-ulps
index e0699c44d8..a6b2f29a6f 100644
--- a/sysdeps/aarch64/libm-test-ulps
+++ b/sysdeps/aarch64/libm-test-ulps
@@ -1248,11 +1248,19 @@ double: 1
 float: 1
 ldouble: 3
 
+Function: "log1p_advsimd":
+double: 1
+float: 1
+
 Function: "log1p_downward":
 double: 1
 float: 2
 ldouble: 3
 
+Function: "log1p_sve":
+double: 1
+float: 1
+
 Function: "log1p_towardzero":
 double: 2
 float: 2
diff --git a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
index 7961a2f374..0f20b5be29 100644
--- a/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
+++ b/sysdeps/unix/sysv/linux/aarch64/libmvec.abilist
@@ -20,6 +20,7 @@ GLIBC_2.39 _ZGVnN2v_atan F
 GLIBC_2.39 _ZGVnN2v_exp10 F
 GLIBC_2.39 _ZGVnN2v_exp2 F
 GLIBC_2.39 _ZGVnN2v_log10 F
+GLIBC_2.39 _ZGVnN2v_log1p F
 GLIBC_2.39 _ZGVnN2v_log2 F
 GLIBC_2.39 _ZGVnN2v_tan F
 GLIBC_2.39 _ZGVnN2vv_atan2 F
@@ -29,6 +30,7 @@ GLIBC_2.39 _ZGVnN4v_atanf F
 GLIBC_2.39 _ZGVnN4v_exp10f F
 GLIBC_2.39 _ZGVnN4v_exp2f F
 GLIBC_2.39 _ZGVnN4v_log10f F
+GLIBC_2.39 _ZGVnN4v_log1pf F
 GLIBC_2.39 _ZGVnN4v_log2f F
 GLIBC_2.39 _ZGVnN4v_tanf F
 GLIBC_2.39 _ZGVnN4vv_atan2f F
@@ -44,6 +46,8 @@ GLIBC_2.39 _ZGVsMxv_exp2 F
 GLIBC_2.39 _ZGVsMxv_exp2f F
 GLIBC_2.39 _ZGVsMxv_log10 F
 GLIBC_2.39 _ZGVsMxv_log10f F
+GLIBC_2.39 _ZGVsMxv_log1p F
+GLIBC_2.39 _ZGVsMxv_log1pf F
 GLIBC_2.39 _ZGVsMxv_log2 F
 GLIBC_2.39 _ZGVsMxv_log2f F
 GLIBC_2.39 _ZGVsMxv_tan F