[PATCH 0/4] Optimized generic expf and exp2f

public inbox for libc-alpha@sourceware.org
 help / color / mirror / Atom feed

* [PATCH 0/4] Optimized generic expf and exp2f
@ 2017-09-13 10:43 Szabolcs Nagy
  2017-09-13 10:45 ` [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers Szabolcs Nagy
                   ` (4 more replies)
  0 siblings, 5 replies; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 10:43 UTC (permalink / raw)
  To: GNU C Library; +Cc: nd

broke down the expf patch into 4 parts.
build-many-glibcs is still running i'll report the
results once it's finished.

Szabolcs Nagy (4):
  Optimized generic expf and exp2f with wrappers
  Move exp compat wrappers under math/
  New expf and exp2f version without SVID compat wrapper
  Do not wrap expf and exp2f

 math/Makefile                                      |   2 +-
 math/Versions                                      |   3 +
 math/w_exp2f.c                                     |   7 +
 math/w_exp2f_compat.c                              |   6 +-
 {sysdeps/ieee754/dbl-64 => math}/w_exp_compat.c    |   0
 math/w_expf.c                                      |   7 +
 {sysdeps/ieee754/flt-32 => math}/w_expf_compat.c   |   7 +-
 math/w_expl_compat.c                               |  48 ++-
 sysdeps/aarch64/fpu/math_private.h                 |  20 ++
 sysdeps/i386/fpu/e_exp2f_data.c                    |   1 +
 sysdeps/i386/fpu/math_errf.c                       |   1 +
 sysdeps/i386/fpu/w_exp2f.c                         |   1 +
 sysdeps/i386/fpu/w_expf.c                          |   1 +
 sysdeps/i386/i686/fpu/multiarch/w_expf.c           |   1 +
 sysdeps/ia64/fpu/e_exp2f_data.c                    |   1 +
 sysdeps/ia64/fpu/math_errf.c                       |   1 +
 sysdeps/ieee754/flt-32/e_exp2f.c                   | 177 ++++-------
 sysdeps/ieee754/flt-32/e_exp2f_data.c              |  44 +++
 sysdeps/ieee754/flt-32/e_expf.c                    | 193 +++++------
 sysdeps/ieee754/flt-32/math_config.h               | 114 +++++++
 sysdeps/ieee754/flt-32/math_errf.c                 |  76 +++++
 sysdeps/ieee754/flt-32/t_exp2f.h                   | 351 ---------------------
 sysdeps/ieee754/flt-32/w_exp2f.c                   |   1 +
 sysdeps/ieee754/flt-32/w_expf.c                    |   1 +
 sysdeps/ieee754/ldbl-128/w_expl_compat.c           |  45 ---
 sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c        |  24 --
 sysdeps/ieee754/ldbl-96/w_expl_compat.c            |  37 ---
 sysdeps/ieee754/ldbl-opt/w_exp_compat.c            |   2 +-
 .../{ldbl-64-128 => ldbl-opt}/w_expl_compat.c      |   2 +-
 sysdeps/m68k/m680x0/fpu/w_expf.c                   |   1 +
 .../powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c |   7 +-
 sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c   |   1 +
 sysdeps/unix/sysv/linux/aarch64/libm.abilist       |   3 +
 sysdeps/unix/sysv/linux/alpha/libm.abilist         |   3 +
 sysdeps/unix/sysv/linux/arm/libm.abilist           |   3 +
 sysdeps/unix/sysv/linux/hppa/libm.abilist          |   3 +
 sysdeps/unix/sysv/linux/i386/libm.abilist          |   3 +
 sysdeps/unix/sysv/linux/ia64/libm.abilist          |   3 +
 sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist |   3 +
 sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist   |   3 +
 sysdeps/unix/sysv/linux/microblaze/libm.abilist    |   3 +
 sysdeps/unix/sysv/linux/mips/mips32/libm.abilist   |   3 +
 sysdeps/unix/sysv/linux/mips/mips64/libm.abilist   |   3 +
 sysdeps/unix/sysv/linux/nios2/libm.abilist         |   3 +
 .../sysv/linux/powerpc/powerpc32/fpu/libm.abilist  |   3 +
 .../linux/powerpc/powerpc32/nofpu/libm.abilist     |   3 +
 .../sysv/linux/powerpc/powerpc64/libm-le.abilist   |   3 +
 .../unix/sysv/linux/powerpc/powerpc64/libm.abilist |   3 +
 sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist  |   3 +
 sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist  |   3 +
 sysdeps/unix/sysv/linux/sh/libm.abilist            |   3 +
 sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist |   3 +
 sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist |   3 +
 .../sysv/linux/tile/tilegx/tilegx32/libm.abilist   |   3 +
 .../sysv/linux/tile/tilegx/tilegx64/libm.abilist   |   3 +
 sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist  |   3 +
 sysdeps/unix/sysv/linux/x86_64/64/libm.abilist     |   3 +
 sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist    |   3 +
 sysdeps/x86_64/fpu/w_expf.c                        |   1 +
 59 files changed, 567 insertions(+), 695 deletions(-)
 create mode 100644 math/w_exp2f.c
 rename {sysdeps/ieee754/dbl-64 => math}/w_exp_compat.c (100%)
 create mode 100644 math/w_expf.c
 rename {sysdeps/ieee754/flt-32 => math}/w_expf_compat.c (88%)
 create mode 100644 sysdeps/i386/fpu/e_exp2f_data.c
 create mode 100644 sysdeps/i386/fpu/math_errf.c
 create mode 100644 sysdeps/i386/fpu/w_exp2f.c
 create mode 100644 sysdeps/i386/fpu/w_expf.c
 create mode 100644 sysdeps/i386/i686/fpu/multiarch/w_expf.c
 create mode 100644 sysdeps/ia64/fpu/e_exp2f_data.c
 create mode 100644 sysdeps/ia64/fpu/math_errf.c
 create mode 100644 sysdeps/ieee754/flt-32/e_exp2f_data.c
 create mode 100644 sysdeps/ieee754/flt-32/math_config.h
 create mode 100644 sysdeps/ieee754/flt-32/math_errf.c
 delete mode 100644 sysdeps/ieee754/flt-32/t_exp2f.h
 create mode 100644 sysdeps/ieee754/flt-32/w_exp2f.c
 create mode 100644 sysdeps/ieee754/flt-32/w_expf.c
 delete mode 100644 sysdeps/ieee754/ldbl-128/w_expl_compat.c
 delete mode 100644 sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c
 delete mode 100644 sysdeps/ieee754/ldbl-96/w_expl_compat.c
 rename sysdeps/ieee754/{ldbl-64-128 => ldbl-opt}/w_expl_compat.c (72%)
 create mode 100644 sysdeps/m68k/m680x0/fpu/w_expf.c
 create mode 100644 sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c
 create mode 100644 sysdeps/x86_64/fpu/w_expf.c

^ permalink raw reply	[flat|nested] 19+ messages in thread

* [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-13 10:43 [PATCH 0/4] Optimized generic expf and exp2f Szabolcs Nagy
@ 2017-09-13 10:45 ` Szabolcs Nagy
  2017-09-13 12:41   ` Joseph Myers
  2017-09-13 20:13   ` Joseph Myers
  2017-09-13 10:46 ` [PATCH 2/4] Move exp compat wrappers under math/ Szabolcs Nagy
                   ` (3 subsequent siblings)
  4 siblings, 2 replies; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 10:45 UTC (permalink / raw)
  To: GNU C Library; +Cc: nd

[-- Attachment #1: Type: text/plain, Size: 2905 bytes --]

Based on new expf and exp2f code from
https://github.com/ARM-software/optimized-routines/

with the new expf benchmark
https://sourceware.org/ml/libc-alpha/2017-08/msg01126.html
aarch64 expf reciprocal-throughput: 3.3x faster
aarch64 expf latency: 1.7x faster
aarch64 libm.so size: -8byte data, -2140byte text
expf/exp2f worst case nearest rounding ulp error: 0.502

Error checks are inline and errno setting is in separate
tail called functions, but the wrappers are kept in this
patch to handle the _LIB_VERSION==_SVID_ case.  (So e.g.
errno is set twice for expf calls and once __expf_finite
calls now on targets where the new code is used.)

Double precision arithmetics is used which is expected
to be faster on most targets (including soft-float) than
using single precision and it is easier to get good
precision result with it.

Const data is kept in a separate translation unit which
complicates maintenance a bit, but is expected to give
good code for literal loads on most targets and allows
sharing data across expf, exp2f and powf. (This data is
disabled on i386 and ia64 which have their own expf, exp2f
and powf code.)

Some configuration is in a new math_config.h the settings
may need further discussion.

Some details may need target specific tweaks:
- best convert and round to int operation in the arg
reduction may be different across targets.
- code was optimized on fma target, optimal polynomial
eval may be different without fma.
- gcc does not always generate good code for fp bit
representation access via unions or it may be inherently
slow on some target.

The libm-test-ulps will need adjustment because..
- The argument reduction ideally uses nearest rounded rint,
but that is not efficient on most targets, so the polynomial
can get evaluated on a wider interval in non-nearest
rounding mode making 1 ulp errors common in that case.
- The polynomial is evaluated such that it has 1 ulp error
on negative tiny inputs with upward rounding, but in
exchange the evaluation is better pipelined.


v2:
- keep the compat wrappers for now.
- cosmetic changes to math_config.h.
- disable math_errf.c and e_exp2f_data.c on i386 and ia64.
- remove unused t_exp2f.h

2017-09-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

	* math/Makefile (type-float-routines): Add math_errf and e_exp2f_data.
	* sysdeps/aarch64/fpu/math_private.h (TOINT_INTRINSICS): Define.
	(roundtoint, converttoint): Likewise.
	* sysdeps/i386/fpu/e_exp2f_data.c: New file.
	* sysdeps/i386/fpu/math_errf.c: New file.
	* sysdeps/ia64/fpu/e_exp2f_data.c: New file.
	* sysdeps/ia64/fpu/math_errf.c: New file.
	* sysdeps/ieee754/flt-32/e_expf.c: New implementation.
	* sysdeps/ieee754/flt-32/e_exp2f.c: New implementation.
	* sysdeps/ieee754/flt-32/e_exp2f_data.c: New file.
	* sysdeps/ieee754/flt-32/math_config.h: New file.
	* sysdeps/ieee754/flt-32/math_errf.c: New file.
	* sysdeps/ieee754/flt-32/t_exp2f.h: Remove.


[-- Attachment #2: 0001-expf.patch --]
[-- Type: text/x-patch, Size: 37687 bytes --]

From 9feaa15d988bcd48a399b8616eb234de805e7b5e Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <szabolcs.nagy@arm.com>
Date: Wed, 6 Sep 2017 17:42:00 +0100
Subject: [PATCH 1/4] expf

---
 math/Makefile                         |   2 +-
 sysdeps/aarch64/fpu/math_private.h    |  20 ++
 sysdeps/i386/fpu/e_exp2f_data.c       |   1 +
 sysdeps/i386/fpu/math_errf.c          |   1 +
 sysdeps/ia64/fpu/e_exp2f_data.c       |   1 +
 sysdeps/ia64/fpu/math_errf.c          |   1 +
 sysdeps/ieee754/flt-32/e_exp2f.c      | 170 ++++++----------
 sysdeps/ieee754/flt-32/e_exp2f_data.c |  44 +++++
 sysdeps/ieee754/flt-32/e_expf.c       | 185 ++++++++----------
 sysdeps/ieee754/flt-32/math_config.h  | 114 +++++++++++
 sysdeps/ieee754/flt-32/math_errf.c    |  76 ++++++++
 sysdeps/ieee754/flt-32/t_exp2f.h      | 351 ----------------------------------
 12 files changed, 399 insertions(+), 567 deletions(-)
 create mode 100644 sysdeps/i386/fpu/e_exp2f_data.c
 create mode 100644 sysdeps/i386/fpu/math_errf.c
 create mode 100644 sysdeps/ia64/fpu/e_exp2f_data.c
 create mode 100644 sysdeps/ia64/fpu/math_errf.c
 create mode 100644 sysdeps/ieee754/flt-32/e_exp2f_data.c
 create mode 100644 sysdeps/ieee754/flt-32/math_config.h
 create mode 100644 sysdeps/ieee754/flt-32/math_errf.c
 delete mode 100644 sysdeps/ieee754/flt-32/t_exp2f.h

diff --git a/math/Makefile b/math/Makefile
index 0601f3ac43..04586156f8 100644
--- a/math/Makefile
+++ b/math/Makefile
@@ -115,7 +115,7 @@ type-double-routines := branred doasin dosincos halfulp mpa mpatan2	\
 
 # float support
 type-float-suffix := f
-type-float-routines := k_rem_pio2f
+type-float-routines := k_rem_pio2f math_errf e_exp2f_data
 
 # _Float128 support
 type-float128-suffix := f128
diff --git a/sysdeps/aarch64/fpu/math_private.h b/sysdeps/aarch64/fpu/math_private.h
index 807111ea5a..775237ffb2 100644
--- a/sysdeps/aarch64/fpu/math_private.h
+++ b/sysdeps/aarch64/fpu/math_private.h
@@ -319,6 +319,26 @@ libc_feresetround_noex_aarch64_ctx (struct rm_ctx *ctx)
 #define libc_feresetround_noexf_ctx	libc_feresetround_noex_aarch64_ctx
 #define libc_feresetround_noexl_ctx	libc_feresetround_noex_aarch64_ctx
 
+/* Hack: only include the large arm_neon.h when needed.  */
+#ifdef _MATH_CONFIG_H
+#include <arm_neon.h>
+
+/* ACLE intrinsics for frintn and fcvtns instructions.  */
+#define TOINT_INTRINSICS 1
+
+static inline double_t
+roundtoint (double_t x)
+{
+  return vget_lane_f64 (vrndn_f64 (vld1_f64 (&x)), 0);
+}
+
+static inline uint64_t
+converttoint (double_t x)
+{
+  return vcvtnd_s64_f64 (x);
+}
+#endif
+
 #include_next <math_private.h>
 
 #endif
diff --git a/sysdeps/i386/fpu/e_exp2f_data.c b/sysdeps/i386/fpu/e_exp2f_data.c
new file mode 100644
index 0000000000..1cc8931700
--- /dev/null
+++ b/sysdeps/i386/fpu/e_exp2f_data.c
@@ -0,0 +1 @@
+/* Not needed.  */
diff --git a/sysdeps/i386/fpu/math_errf.c b/sysdeps/i386/fpu/math_errf.c
new file mode 100644
index 0000000000..1cc8931700
--- /dev/null
+++ b/sysdeps/i386/fpu/math_errf.c
@@ -0,0 +1 @@
+/* Not needed.  */
diff --git a/sysdeps/ia64/fpu/e_exp2f_data.c b/sysdeps/ia64/fpu/e_exp2f_data.c
new file mode 100644
index 0000000000..1cc8931700
--- /dev/null
+++ b/sysdeps/ia64/fpu/e_exp2f_data.c
@@ -0,0 +1 @@
+/* Not needed.  */
diff --git a/sysdeps/ia64/fpu/math_errf.c b/sysdeps/ia64/fpu/math_errf.c
new file mode 100644
index 0000000000..1cc8931700
--- /dev/null
+++ b/sysdeps/ia64/fpu/math_errf.c
@@ -0,0 +1 @@
+/* Not needed.  */
diff --git a/sysdeps/ieee754/flt-32/e_exp2f.c b/sysdeps/ieee754/flt-32/e_exp2f.c
index 567d3ff6d0..be53e5c699 100644
--- a/sysdeps/ieee754/flt-32/e_exp2f.c
+++ b/sysdeps/ieee754/flt-32/e_exp2f.c
@@ -1,7 +1,6 @@
-/* Single-precision floating point 2^x.
-   Copyright (C) 1997-2017 Free Software Foundation, Inc.
+/* Single-precision 2^x function.
+   Copyright (C) 2017 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
-   Contributed by Geoffrey Keating <geoffk@ozemail.com.au>
 
    The GNU C Library is free software; you can redistribute it and/or
    modify it under the terms of the GNU Lesser General Public
@@ -17,116 +16,73 @@
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-/* The basic design here is from
-   Shmuel Gal and Boris Bachelis, "An Accurate Elementary Mathematical
-   Library for the IEEE Floating Point Standard", ACM Trans. Math. Soft.,
-   17 (1), March 1991, pp. 26-45.
-   It has been slightly modified to compute 2^x instead of e^x, and for
-   single-precision.
-   */
-#ifndef _GNU_SOURCE
-# define _GNU_SOURCE
-#endif
-#include <stdlib.h>
-#include <float.h>
-#include <ieee754.h>
 #include <math.h>
-#include <fenv.h>
-#include <inttypes.h>
-#include <math_private.h>
-
-#include "t_exp2f.h"
-
-static const float TWOM100 = 7.88860905e-31;
-static const float TWO127 = 1.7014118346e+38;
+#include <stdint.h>
+#include "math_config.h"
+
+/*
+EXP2F_TABLE_BITS = 5
+EXP2F_POLY_ORDER = 3
+
+ULP error: 0.502 (nearest rounding.)
+Relative error: 1.69 * 2^-34 in [-1/64, 1/64] (before rounding.)
+Wrong count: 168353 (all nearest rounding wrong results with fma.)
+Non-nearest ULP error: 1 (rounded ULP error)
+*/
+
+#define N (1 << EXP2F_TABLE_BITS)
+#define T __exp2f_data.tab
+#define C __exp2f_data.poly
+#define SHIFT __exp2f_data.shift_scaled
+
+static inline uint32_t
+top12 (float x)
+{
+  return asuint (x) >> 20;
+}
 
 float
 __ieee754_exp2f (float x)
 {
-  static const float himark = (float) FLT_MAX_EXP;
-  static const float lomark = (float) (FLT_MIN_EXP - FLT_MANT_DIG - 1);
-
-  /* Check for usual case.  */
-  if (isless (x, himark) && isgreaterequal (x, lomark))
+  uint32_t abstop;
+  uint64_t ki, t;
+  /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
+  double_t kd, xd, z, r, r2, y, s;
+
+  xd = (double_t) x;
+  abstop = top12 (x) & 0x7ff;
+  if (__glibc_unlikely (abstop >= top12 (128.0f)))
     {
-      static const float THREEp14 = 49152.0;
-      int tval, unsafe;
-      float rx, x22, result;
-      union ieee754_float ex2_u, scale_u;
-
-      if (fabsf (x) < FLT_EPSILON / 4.0f)
-	return 1.0f + x;
-
-      {
-	SET_RESTORE_ROUND_NOEXF (FE_TONEAREST);
-
-	/* 1. Argument reduction.
-	   Choose integers ex, -128 <= t < 128, and some real
-	   -1/512 <= x1 <= 1/512 so that
-	   x = ex + t/512 + x1.
-
-	   First, calculate rx = ex + t/256.  */
-	rx = x + THREEp14;
-	rx -= THREEp14;
-	x -= rx;  /* Compute x=x1. */
-	/* Compute tval = (ex*256 + t)+128.
-	   Now, t = (tval mod 256)-128 and ex=tval/256  [that's mod, NOT %;
-	   and /-round-to-nearest not the usual c integer /].  */
-	tval = (int) (rx * 256.0f + 128.0f);
-
-	/* 2. Adjust for accurate table entry.
-	   Find e so that
-	   x = ex + t/256 + e + x2
-	   where -7e-4 < e < 7e-4, and
-	   (float)(2^(t/256+e))
-	   is accurate to one part in 2^-64.  */
-
-	/* 'tval & 255' is the same as 'tval%256' except that it's always
-	   positive.
-	   Compute x = x2.  */
-	x -= __exp2f_deltatable[tval & 255];
-
-	/* 3. Compute ex2 = 2^(t/255+e+ex).  */
-	ex2_u.f = __exp2f_atable[tval & 255];
-	tval >>= 8;
-	/* x2 is an integer multiple of 2^-30; avoid intermediate
-	   underflow from the calculation of x22 * x.  */
-	unsafe = abs(tval) >= -FLT_MIN_EXP - 32;
-	ex2_u.ieee.exponent += tval >> unsafe;
-	scale_u.f = 1.0;
-	scale_u.ieee.exponent += tval - (tval >> unsafe);
-
-	/* 4. Approximate 2^x2 - 1, using a second-degree polynomial,
-	   with maximum error in [-2^-9 - 2^-14, 2^-9 + 2^-14]
-	   less than 1.3e-10.  */
-
-	x22 = (.24022656679f * x + .69314736128f) * ex2_u.f;
-      }
-
-      /* 5. Return (2^x2-1) * 2^(t/512+e+ex) + 2^(t/512+e+ex).  */
-      result = x22 * x + ex2_u.f;
-
-      if (!unsafe)
-	return result;
-      else
-	{
-	  result *= scale_u.f;
-	  math_check_force_underflow_nonneg (result);
-	  return result;
-	}
-    }
-  /* Exceptional cases:  */
-  else if (isless (x, himark))
-    {
-      if (isinf (x))
-	/* e^-inf == 0, with no error.  */
-	return 0;
-      else
-	/* Underflow */
-	return TWOM100 * TWOM100;
+      /* |x| >= 128 or x is nan.  */
+      if (asuint (x) == asuint (-INFINITY))
+	return 0.0f;
+      if (abstop >= top12 (INFINITY))
+	return x + x;
+      if (x > 0.0f)
+	return __math_oflowf (0);
+      if (x <= -150.0f)
+	return __math_uflowf (0);
+#if WANT_ERRNO_UFLOW
+      if (x < -149.0f)
+	return __math_may_uflowf (0);
+#endif
     }
-  else
-    /* Return x, if x is a NaN or Inf; or overflow, otherwise.  */
-    return TWO127*x;
+
+  /* x = k/N + r with r in [-1/(2N), 1/(2N)] and int k.  */
+  kd = (double) (xd + SHIFT); /* Rounding to double precision is required.  */
+  ki = asuint64 (kd);
+  kd -= SHIFT; /* k/N for int k.  */
+  r = xd - kd;
+
+  /* exp2(x) = 2^(k/N) * 2^r ~= s * (C0*r^3 + C1*r^2 + C2*r + 1) */
+  t = T[ki % N];
+  t += ki << (52 - EXP2F_TABLE_BITS);
+  s = asdouble (t);
+  z = C[0] * r + C[1];
+  r2 = r * r;
+  y = C[2] * r + 1;
+  y = z * r2 + y;
+  y = y * s;
+  return (float) y;
 }
 strong_alias (__ieee754_exp2f, __exp2f_finite)
diff --git a/sysdeps/ieee754/flt-32/e_exp2f_data.c b/sysdeps/ieee754/flt-32/e_exp2f_data.c
new file mode 100644
index 0000000000..390dcae333
--- /dev/null
+++ b/sysdeps/ieee754/flt-32/e_exp2f_data.c
@@ -0,0 +1,44 @@
+/* Shared data between expf, exp2f and powf.
+   Copyright (C) 2017 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include "math_config.h"
+
+#define N (1 << EXP2F_TABLE_BITS)
+
+const struct exp2f_data __exp2f_data = {
+  /* tab[i] = uint(2^(i/N)) - (i << 52-BITS)
+     used for computing 2^(k/N) for an int |k| < 150 N as
+     double(tab[k%N] + (k << 52-BITS)) */
+  .tab = {
+0x3ff0000000000000, 0x3fefd9b0d3158574, 0x3fefb5586cf9890f, 0x3fef9301d0125b51,
+0x3fef72b83c7d517b, 0x3fef54873168b9aa, 0x3fef387a6e756238, 0x3fef1e9df51fdee1,
+0x3fef06fe0a31b715, 0x3feef1a7373aa9cb, 0x3feedea64c123422, 0x3feece086061892d,
+0x3feebfdad5362a27, 0x3feeb42b569d4f82, 0x3feeab07dd485429, 0x3feea47eb03a5585,
+0x3feea09e667f3bcd, 0x3fee9f75e8ec5f74, 0x3feea11473eb0187, 0x3feea589994cce13,
+0x3feeace5422aa0db, 0x3feeb737b0cdc5e5, 0x3feec49182a3f090, 0x3feed503b23e255d,
+0x3feee89f995ad3ad, 0x3feeff76f2fb5e47, 0x3fef199bdd85529c, 0x3fef3720dcef9069,
+0x3fef5818dcfba487, 0x3fef7c97337b9b5f, 0x3fefa4afa2a490da, 0x3fefd0765b6e4540,
+  },
+  .shift_scaled = 0x1.8p+52 / N,
+  .poly = { 0x1.c6af84b912394p-5, 0x1.ebfce50fac4f3p-3, 0x1.62e42ff0c52d6p-1 },
+  .shift = 0x1.8p+52,
+  .invln2_scaled = 0x1.71547652b82fep+0 * N,
+  .poly_scaled = {
+0x1.c6af84b912394p-5/N/N/N, 0x1.ebfce50fac4f3p-3/N/N, 0x1.62e42ff0c52d6p-1/N
+  },
+};
diff --git a/sysdeps/ieee754/flt-32/e_expf.c b/sysdeps/ieee754/flt-32/e_expf.c
index 782072f213..b8244a82cf 100644
--- a/sysdeps/ieee754/flt-32/e_expf.c
+++ b/sysdeps/ieee754/flt-32/e_expf.c
@@ -1,7 +1,6 @@
-/* Single-precision floating point e^x.
-   Copyright (C) 1997-2017 Free Software Foundation, Inc.
+/* Single-precision e^x function.
+   Copyright (C) 2017 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
-   Contributed by Geoffrey Keating <geoffk@ozemail.com.au>
 
    The GNU C Library is free software; you can redistribute it and/or
    modify it under the terms of the GNU Lesser General Public
@@ -17,117 +16,87 @@
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
-/* How this works:
-
-   The input value, x, is written as
-
-   x = n * ln(2) + t/512 + delta[t] + x;
-
-   where:
-   - n is an integer, 127 >= n >= -150;
-   - t is an integer, 177 >= t >= -177
-   - delta is based on a table entry, delta[t] < 2^-28
-   - x is whatever is left, |x| < 2^-10
-
-   Then e^x is approximated as
-
-   e^x = 2^n ( e^(t/512 + delta[t])
-	       + ( e^(t/512 + delta[t])
-		   * ( p(x + delta[t] + n * ln(2)) - delta ) ) )
-
-   where
-   - p(x) is a polynomial approximating e(x)-1;
-   - e^(t/512 + delta[t]) is obtained from a table.
-
-   The table used is the same one as for the double precision version;
-   since we have the table, we might as well use it.
-
-   It turns out to be faster to do calculations in double precision than
-   to perform an 'accurate table method' expf, because of the range reduction
-   overhead (compare exp2f).
-   */
-#include <float.h>
-#include <ieee754.h>
 #include <math.h>
-#include <fenv.h>
-#include <inttypes.h>
-#include <math_private.h>
-
-extern const float __exp_deltatable[178];
-extern const double __exp_atable[355] /* __attribute__((mode(DF))) */;
-
-static const float TWOM100 = 7.88860905e-31;
-static const float TWO127 = 1.7014118346e+38;
+#include <stdint.h>
+#include "math_config.h"
+
+/*
+EXP2F_TABLE_BITS = 5
+EXP2F_POLY_ORDER = 3
+
+ULP error: 0.502 (nearest rounding.)
+Relative error: 1.69 * 2^-34 in [-ln2/64, ln2/64] (before rounding.)
+Wrong count: 170635 (all nearest rounding wrong results with fma.)
+Non-nearest ULP error: 1 (rounded ULP error)
+*/
+
+#define N (1 << EXP2F_TABLE_BITS)
+#define InvLn2N __exp2f_data.invln2_scaled
+#define T __exp2f_data.tab
+#define C __exp2f_data.poly_scaled
+
+static inline uint32_t
+top12 (float x)
+{
+  return asuint (x) >> 20;
+}
 
 float
 __ieee754_expf (float x)
 {
-  static const float himark = 88.72283935546875;
-  static const float lomark = -103.972084045410;
-  /* Check for usual case.  */
-  if (isless (x, himark) && isgreater (x, lomark))
+  uint32_t abstop;
+  uint64_t ki, t;
+  /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
+  double_t kd, xd, z, r, r2, y, s;
+
+  xd = (double_t) x;
+  abstop = top12 (x) & 0x7ff;
+  if (__glibc_unlikely (abstop >= top12 (88.0f)))
     {
-      static const float THREEp42 = 13194139533312.0;
-      static const float THREEp22 = 12582912.0;
-      /* 1/ln(2).  */
-#undef M_1_LN2
-      static const float M_1_LN2 = 1.44269502163f;
-      /* ln(2) */
-#undef M_LN2
-      static const double M_LN2 = .6931471805599452862;
-
-      int tval;
-      double x22, t, result, dx;
-      float n, delta;
-      union ieee754_double ex2_u;
-
-      {
-	SET_RESTORE_ROUND_NOEXF (FE_TONEAREST);
-
-	/* Calculate n.  */
-	n = x * M_1_LN2 + THREEp22;
-	n -= THREEp22;
-	dx = x - n*M_LN2;
-
-	/* Calculate t/512.  */
-	t = dx + THREEp42;
-	t -= THREEp42;
-	dx -= t;
-
-	/* Compute tval = t.  */
-	tval = (int) (t * 512.0);
-
-	if (t >= 0)
-	  delta = - __exp_deltatable[tval];
-	else
-	  delta = __exp_deltatable[-tval];
-
-	/* Compute ex2 = 2^n e^(t/512+delta[t]).  */
-	ex2_u.d = __exp_atable[tval+177];
-	ex2_u.ieee.exponent += (int) n;
-
-	/* Approximate e^(dx+delta) - 1, using a second-degree polynomial,
-	   with maximum error in [-2^-10-2^-28,2^-10+2^-28]
-	   less than 5e-11.  */
-	x22 = (0.5000000496709180453 * dx + 1.0000001192102037084) * dx + delta;
-      }
-
-      /* Return result.  */
-      result = x22 * ex2_u.d + ex2_u.d;
-      return (float) result;
+      /* |x| >= 88 or x is nan.  */
+      if (asuint (x) == asuint (-INFINITY))
+	return 0.0f;
+      if (abstop >= top12 (INFINITY))
+	return x + x;
+      if (x > 0x1.62e42ep6f) /* x > log(0x1p128) ~= 88.72 */
+	return __math_oflowf (0);
+      if (x < -0x1.9fe368p6f) /* x < log(0x1p-150) ~= -103.97 */
+	return __math_uflowf (0);
+#if WANT_ERRNO_UFLOW
+      if (x < -0x1.9d1d9ep6f) /* x < log(0x1p-149) ~= -103.28 */
+	return __math_may_uflowf (0);
+#endif
     }
-  /* Exceptional cases:  */
-  else if (isless (x, himark))
-    {
-      if (isinf (x))
-	/* e^-inf == 0, with no error.  */
-	return 0;
-      else
-	/* Underflow */
-	return TWOM100 * TWOM100;
-    }
-  else
-    /* Return x, if x is a NaN or Inf; or overflow, otherwise.  */
-    return TWO127*x;
+
+  /* x*N/Ln2 = k + r with r in [-1/2, 1/2] and int k.  */
+  z = InvLn2N * xd;
+
+  /* Round and convert z to int, the result is in [-150*N, 128*N] and
+     ideally ties-to-even rule is used, otherwise the magnitude of r
+     can be bigger which gives larger approximation error.  */
+#if TOINT_INTRINSICS
+  kd = roundtoint (z);
+  ki = converttoint (z);
+#elif TOINT_RINT
+  kd = rint (z);
+  ki = (long) kd;
+#elif TOINT_SHIFT
+# define SHIFT __exp2f_data.shift
+  kd = (double) (z + SHIFT); /* Rounding to double precision is required.  */
+  ki = asuint64 (kd);
+  kd -= SHIFT;
+#endif
+  r = z - kd;
+
+  /* exp(x) = 2^(k/N) * 2^(r/N) ~= s * (C0*r^3 + C1*r^2 + C2*r + 1) */
+  t = T[ki % N];
+  t += ki << (52 - EXP2F_TABLE_BITS);
+  s = asdouble (t);
+  z = C[0] * r + C[1];
+  r2 = r * r;
+  y = C[2] * r + 1;
+  y = z * r2 + y;
+  y = y * s;
+  return (float) y;
 }
 strong_alias (__ieee754_expf, __expf_finite)
diff --git a/sysdeps/ieee754/flt-32/math_config.h b/sysdeps/ieee754/flt-32/math_config.h
new file mode 100644
index 0000000000..1759c447c8
--- /dev/null
+++ b/sysdeps/ieee754/flt-32/math_config.h
@@ -0,0 +1,114 @@
+/* Configuration for math routines.
+   Copyright (C) 2017 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
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _MATH_CONFIG_H
+#define _MATH_CONFIG_H
+
+#include <math.h>
+#include <math_private.h>
+#include <stdint.h>
+
+#ifndef WANT_ROUNDING
+/* Correct special case results in non-nearest rounding modes.  */
+# define WANT_ROUNDING 1
+#endif
+#ifndef WANT_ERRNO
+/* Set errno according to ISO C with (math_errhandling & MATH_ERRNO) != 0.  */
+# define WANT_ERRNO 1
+#endif
+#ifndef WANT_ERRNO_UFLOW
+/* Set errno to ERANGE if result underflows to 0 (in all rounding modes).  */
+# define WANT_ERRNO_UFLOW (WANT_ROUNDING && WANT_ERRNO)
+#endif
+
+#ifndef TOINT_INTRINSICS
+# define TOINT_INTRINSICS 0
+#endif
+#ifndef TOINT_RINT
+# define TOINT_RINT 0
+#endif
+#ifndef TOINT_SHIFT
+# define TOINT_SHIFT 1
+#endif
+
+static inline uint32_t
+asuint (float f)
+{
+  union
+  {
+    float f;
+    uint32_t i;
+  } u = {f};
+  return u.i;
+}
+
+static inline float
+asfloat (uint32_t i)
+{
+  union
+  {
+    uint32_t i;
+    float f;
+  } u = {i};
+  return u.f;
+}
+
+static inline uint64_t
+asuint64 (double f)
+{
+  union
+  {
+    double f;
+    uint64_t i;
+  } u = {f};
+  return u.i;
+}
+
+static inline double
+asdouble (uint64_t i)
+{
+  union
+  {
+    uint64_t i;
+    double f;
+  } u = {i};
+  return u.f;
+}
+
+#define NOINLINE __attribute__ ((noinline))
+
+attribute_hidden float __math_oflowf (unsigned long);
+attribute_hidden float __math_uflowf (unsigned long);
+attribute_hidden float __math_may_uflowf (unsigned long);
+attribute_hidden float __math_divzerof (unsigned long);
+attribute_hidden float __math_invalidf (float);
+
+/* Shared between expf, exp2f and powf.  */
+#define EXP2F_TABLE_BITS 5
+#define EXP2F_POLY_ORDER 3
+extern const struct exp2f_data
+{
+  uint64_t tab[1 << EXP2F_TABLE_BITS];
+  double shift_scaled;
+  double poly[EXP2F_POLY_ORDER];
+  double shift;
+  double invln2_scaled;
+  double poly_scaled[EXP2F_POLY_ORDER];
+} __exp2f_data;
+
+#endif
diff --git a/sysdeps/ieee754/flt-32/math_errf.c b/sysdeps/ieee754/flt-32/math_errf.c
new file mode 100644
index 0000000000..ab546e24cf
--- /dev/null
+++ b/sysdeps/ieee754/flt-32/math_errf.c
@@ -0,0 +1,76 @@
+/* Single-precision math error handling.
+   Copyright (C) 2017 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include "math_config.h"
+
+#if WANT_ERRNO
+# include <errno.h>
+/* NOINLINE reduces code size.  */
+NOINLINE static float
+with_errnof (float y, int e)
+{
+  errno = e;
+  return y;
+}
+#else
+# define with_errnof(x, e) (x)
+#endif
+
+/* NOINLINE prevents fenv semantics breaking optimizations.  */
+NOINLINE static float
+xflowf (unsigned long sign, float y)
+{
+  y = (sign ? -y : y) * y;
+  return with_errnof (y, ERANGE);
+}
+
+attribute_hidden float
+__math_uflowf (unsigned long sign)
+{
+  return xflowf (sign, 0x1p-95f);
+}
+
+#if WANT_ERRNO_UFLOW
+/* Underflows to zero in some non-nearest rounding mode, setting errno
+   is valid even if the result is non-zero, but in the subnormal range.  */
+attribute_hidden float
+__math_may_uflowf (unsigned long sign)
+{
+  return xflowf (sign, 0x1.4p-75f);
+}
+#endif
+
+attribute_hidden float
+__math_oflowf (unsigned long sign)
+{
+  return xflowf (sign, 0x1p97f);
+}
+
+attribute_hidden float
+__math_divzerof (unsigned long sign)
+{
+  float y = 0;
+  return with_errnof ((sign ? -1 : 1) / y, ERANGE);
+}
+
+attribute_hidden float
+__math_invalidf (float x)
+{
+  float y = (x - x) / (x - x);
+  return isnan (x) ? y : with_errnof (y, EDOM);
+}
diff --git a/sysdeps/ieee754/flt-32/t_exp2f.h b/sysdeps/ieee754/flt-32/t_exp2f.h
deleted file mode 100644
index aecabcc372..0000000000
--- a/sysdeps/ieee754/flt-32/t_exp2f.h
+++ /dev/null
@@ -1,351 +0,0 @@
-/* Accurate tables for exp2f().
-   Copyright (C) 1998-2017 Free Software Foundation, Inc.
-   This file is part of the GNU C Library.
-   Contributed by Geoffrey Keating <geoffk@ozemail.com.au>
-
-   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
-   <http://www.gnu.org/licenses/>.  */
-
-/* This table has the property that, for all integers -128 <= i <= 127,
-   exp(i/256.0 + __exp2f_deltatable[i-128]) == __exp2f_atable[i+128] + r
-   for some -2^-35 < r < 2^-35 (abs(r) < 2^-36 if i <= 0); and that
-   __exp2f_deltatable[i+128] == t * 2^-30
-   for integer t so that abs(t) <= 43447 * 2^0.  */
-
-#define W30 (9.31322575e-10)
-static const float __exp2f_deltatable[256] = {
-      -810*W30,       283*W30,     -1514*W30,      1304*W30,
-     -1148*W30,       -98*W30,      -744*W30,      -156*W30,
-      -419*W30,      -155*W30,       474*W30,       167*W30,
-     -1984*W30,      -826*W30,       692*W30,       781*W30,
-      -578*W30,      -411*W30,      -129*W30,     -1500*W30,
-       654*W30,      -141*W30,      -816*W30,       -53*W30,
-       148*W30,       493*W30,     -2214*W30,       760*W30,
-       260*W30,       750*W30,     -1300*W30,      1424*W30,
-     -1445*W30,      -339*W30,      -680*W30,      -349*W30,
-      -922*W30,       531*W30,       193*W30,     -2892*W30,
-       290*W30,     -2145*W30,      -276*W30,       485*W30,
-      -695*W30,       215*W30,     -7093*W30,       412*W30,
-     -4596*W30,       367*W30,       592*W30,      -615*W30,
-       -97*W30,     -1066*W30,       972*W30,      -226*W30,
-      -625*W30,      -374*W30,     -5647*W30,      -180*W30,
-     20349*W30,      -447*W30,       111*W30,     -4164*W30,
-       -87*W30,       -21*W30,      -251*W30,        66*W30,
-      -517*W30,      2093*W30,      -263*W30,       182*W30,
-      -601*W30,       475*W30,      -483*W30,     -1251*W30,
-      -373*W30,      1471*W30,       -92*W30,      -215*W30,
-       -97*W30,      -190*W30,         0*W30,      -290*W30,
-     -2647*W30,      1940*W30,      -582*W30,        28*W30,
-       833*W30,      1493*W30,        34*W30,       321*W30,
-      3327*W30,       -35*W30,       177*W30,      -135*W30,
-      -796*W30,      -428*W30,       129*W30,      9332*W30,
-       -12*W30,       -69*W30,     -1743*W30,      6508*W30,
-       -60*W30,       359*W30,     43447*W30,        15*W30,
-       -23*W30,      -305*W30,      -375*W30,      -652*W30,
-       667*W30,       269*W30,     -1575*W30,       185*W30,
-      -329*W30,       200*W30,      6002*W30,       163*W30,
-      -647*W30,        19*W30,      -603*W30,      -755*W30,
-       742*W30,      -438*W30,      3587*W30,      2560*W30,
-         0*W30,      -520*W30,      -241*W30,      -299*W30,
-     -1270*W30,      -991*W30,     -1138*W30,       255*W30,
-     -1192*W30,      1722*W30,      1023*W30,      3700*W30,
-     -1388*W30,     -1551*W30,     -2549*W30,        27*W30,
-       282*W30,       673*W30,       113*W30,      1561*W30,
-        72*W30,       873*W30,        87*W30,      -395*W30,
-      -433*W30,       629*W30,      3440*W30,      -284*W30,
-      -592*W30,      -103*W30,       -46*W30,     -3844*W30,
-      1712*W30,       303*W30,      1555*W30,      -631*W30,
-     -1400*W30,      -961*W30,      -854*W30,      -276*W30,
-       407*W30,       833*W30,      -345*W30,     -1501*W30,
-       121*W30,     -1581*W30,       400*W30,       150*W30,
-      1224*W30,      -139*W30,      -563*W30,       879*W30,
-       933*W30,      2939*W30,       788*W30,       211*W30,
-       530*W30,      -192*W30,       706*W30,    -13347*W30,
-      1065*W30,         3*W30,       111*W30,      -208*W30,
-      -360*W30,      -532*W30,      -291*W30,       483*W30,
-       987*W30,       -33*W30,     -1373*W30,      -166*W30,
-     -1174*W30,     -3955*W30,      1601*W30,      -280*W30,
-      1405*W30,       600*W30,     -1659*W30,       -23*W30,
-       390*W30,       449*W30,       570*W30,    -13143*W30,
-        -9*W30,     -1646*W30,      1201*W30,       294*W30,
-      2181*W30,     -1173*W30,      1388*W30,     -4504*W30,
-       190*W30,     -2304*W30,       211*W30,       239*W30,
-        48*W30,      -817*W30,      1018*W30,      1828*W30,
-      -663*W30,      1408*W30,       408*W30,       -36*W30,
-      1295*W30,      -230*W30,      1341*W30,         9*W30,
-        40*W30,       705*W30,       186*W30,       376*W30,
-       557*W30,      5866*W30,       363*W30,     -1558*W30,
-       718*W30,       669*W30,      1369*W30,     -2972*W30,
-      -468*W30,      -121*W30,      -219*W30,       667*W30,
-     29954*W30,       366*W30,        48*W30,      -203*W30
-};
-
-static const float __exp2f_atable[256] /* __attribute__((mode(SF))) */ = {
- 0.707106411447, /* 0x0.b504ecfff */
- 0.709024071690, /* 0x0.b58299fff */
- 0.710945606239, /* 0x0.b60088000 */
- 0.712874472142, /* 0x0.b67ef1000 */
- 0.714806139464, /* 0x0.b6fd88fff */
- 0.716744661340, /* 0x0.b77c94000 */
- 0.718687653549, /* 0x0.b7fbea000 */
- 0.720636486992, /* 0x0.b87ba1fff */
- 0.722590208040, /* 0x0.b8fbabfff */
- 0.724549472323, /* 0x0.b97c12fff */
- 0.726514220228, /* 0x0.b9fcd5fff */
- 0.728483855735, /* 0x0.ba7deb000 */
- 0.730457961549, /* 0x0.baff4afff */
- 0.732438981522, /* 0x0.bb811efff */
- 0.734425544748, /* 0x0.bc0350000 */
- 0.736416816713, /* 0x0.bc85d0000 */
- 0.738412797450, /* 0x0.bd089efff */
- 0.740414917465, /* 0x0.bd8bd4fff */
- 0.742422521111, /* 0x0.be0f66fff */
- 0.744434773914, /* 0x0.be9346fff */
- 0.746454179287, /* 0x0.bf179f000 */
- 0.748477637755, /* 0x0.bf9c3afff */
- 0.750506639473, /* 0x0.c02133fff */
- 0.752541840064, /* 0x0.c0a694fff */
- 0.754582285889, /* 0x0.c12c4e000 */
- 0.756628334525, /* 0x0.c1b265000 */
- 0.758678436269, /* 0x0.c238bffff */
- 0.760736882681, /* 0x0.c2bfa6fff */
- 0.762799203401, /* 0x0.c346cf000 */
- 0.764867603790, /* 0x0.c3ce5d000 */
- 0.766940355298, /* 0x0.c45633fff */
- 0.769021093841, /* 0x0.c4de90fff */
- 0.771104693409, /* 0x0.c5671dfff */
- 0.773195922364, /* 0x0.c5f02afff */
- 0.775292098512, /* 0x0.c6798afff */
- 0.777394294745, /* 0x0.c70350000 */
- 0.779501736166, /* 0x0.c78d6d000 */
- 0.781615912910, /* 0x0.c817fafff */
- 0.783734917628, /* 0x0.c8a2d9fff */
- 0.785858273516, /* 0x0.c92e02000 */
- 0.787990570071, /* 0x0.c9b9c0000 */
- 0.790125787245, /* 0x0.ca45aefff */
- 0.792268991467, /* 0x0.cad223fff */
- 0.794417440881, /* 0x0.cb5ef0fff */
- 0.796570718287, /* 0x0.cbec0efff */
- 0.798730909811, /* 0x0.cc79a0fff */
- 0.800892710672, /* 0x0.cd074dfff */
- 0.803068041795, /* 0x0.cd95ddfff */
- 0.805242776881, /* 0x0.ce2464000 */
- 0.807428598393, /* 0x0.ceb3a3fff */
- 0.809617877002, /* 0x0.cf431dfff */
- 0.811812341211, /* 0x0.cfd2eefff */
- 0.814013659956, /* 0x0.d06333000 */
- 0.816220164311, /* 0x0.d0f3ce000 */
- 0.818434238424, /* 0x0.d184e7fff */
- 0.820652604094, /* 0x0.d21649fff */
- 0.822877407074, /* 0x0.d2a818000 */
- 0.825108587751, /* 0x0.d33a51000 */
- 0.827342867839, /* 0x0.d3ccbdfff */
- 0.829588949684, /* 0x0.d45ff1000 */
- 0.831849217401, /* 0x0.d4f411fff */
- 0.834093391880, /* 0x0.d58724fff */
- 0.836355149750, /* 0x0.d61b5f000 */
- 0.838620424257, /* 0x0.d6afd3fff */
- 0.840896368027, /* 0x0.d744fc000 */
- 0.843176305293, /* 0x0.d7da66fff */
- 0.845462262643, /* 0x0.d87037000 */
- 0.847754716864, /* 0x0.d90673fff */
- 0.850052893157, /* 0x0.d99d10fff */
- 0.852359056469, /* 0x0.da3433fff */
- 0.854668736446, /* 0x0.dacb91fff */
- 0.856986224651, /* 0x0.db6373000 */
- 0.859309315673, /* 0x0.dbfbb1fff */
- 0.861639738080, /* 0x0.dc946bfff */
- 0.863975346095, /* 0x0.dd2d7d000 */
- 0.866317391394, /* 0x0.ddc6f9fff */
- 0.868666708472, /* 0x0.de60f1000 */
- 0.871022939695, /* 0x0.defb5c000 */
- 0.873383641229, /* 0x0.df9611fff */
- 0.875751554968, /* 0x0.e03141000 */
- 0.878126025200, /* 0x0.e0ccde000 */
- 0.880506813521, /* 0x0.e168e4fff */
- 0.882894217966, /* 0x0.e2055afff */
- 0.885287821299, /* 0x0.e2a239000 */
- 0.887686729423, /* 0x0.e33f6ffff */
- 0.890096127973, /* 0x0.e3dd56fff */
- 0.892507970338, /* 0x0.e47b67000 */
- 0.894928157336, /* 0x0.e51a03000 */
- 0.897355020043, /* 0x0.e5b90efff */
- 0.899788379682, /* 0x0.e65888000 */
- 0.902227103705, /* 0x0.e6f85afff */
- 0.904673457151, /* 0x0.e798ae000 */
- 0.907128036008, /* 0x0.e8398afff */
- 0.909585535528, /* 0x0.e8da99000 */
- 0.912051796915, /* 0x0.e97c3a000 */
- 0.914524436003, /* 0x0.ea1e46000 */
- 0.917003571999, /* 0x0.eac0bf000 */
- 0.919490039339, /* 0x0.eb63b2fff */
- 0.921983361257, /* 0x0.ec071a000 */
- 0.924488604054, /* 0x0.ecab48fff */
- 0.926989555360, /* 0x0.ed4f30000 */
- 0.929502844812, /* 0x0.edf3e6000 */
- 0.932021975503, /* 0x0.ee98fdfff */
- 0.934553921208, /* 0x0.ef3eecfff */
- 0.937083780759, /* 0x0.efe4b8fff */
- 0.939624726786, /* 0x0.f08b3f000 */
- 0.942198514924, /* 0x0.f133ebfff */
- 0.944726586343, /* 0x0.f1d99a000 */
- 0.947287976728, /* 0x0.f28176fff */
- 0.949856162070, /* 0x0.f329c5fff */
- 0.952431440345, /* 0x0.f3d28bfff */
- 0.955013573175, /* 0x0.f47bc5000 */
- 0.957603693021, /* 0x0.f52584000 */
- 0.960199773321, /* 0x0.f5cfa7000 */
- 0.962801992906, /* 0x0.f67a31000 */
- 0.965413510788, /* 0x0.f72556fff */
- 0.968030691152, /* 0x0.f7d0dc000 */
- 0.970655620084, /* 0x0.f87ce2fff */
- 0.973290979849, /* 0x0.f92998fff */
- 0.975926160805, /* 0x0.f9d64bfff */
- 0.978571653370, /* 0x0.fa83ac000 */
- 0.981225252139, /* 0x0.fb3193fff */
- 0.983885228626, /* 0x0.fbdfe6fff */
- 0.986552715296, /* 0x0.fc8eb7fff */
- 0.989228487027, /* 0x0.fd3e14000 */
- 0.991909801964, /* 0x0.fdedcd000 */
- 0.994601726545, /* 0x0.fe9e38000 */
- 0.997297704209, /* 0x0.ff4ee6fff */
- 1.000000000000, /* 0x1.000000000 */
- 1.002710938457, /* 0x1.00b1aa000 */
- 1.005429744692, /* 0x1.0163d7ffe */
- 1.008155703526, /* 0x1.02167dffe */
- 1.010888457284, /* 0x1.02c995fff */
- 1.013629436498, /* 0x1.037d38000 */
- 1.016377568250, /* 0x1.043152000 */
- 1.019134163841, /* 0x1.04e5f9ffe */
- 1.021896362316, /* 0x1.059b00000 */
- 1.024668931945, /* 0x1.0650b3ffe */
- 1.027446627635, /* 0x1.0706be001 */
- 1.030234098408, /* 0x1.07bd6bffe */
- 1.033023953416, /* 0x1.087441ffe */
- 1.035824656494, /* 0x1.092bce000 */
- 1.038632392900, /* 0x1.09e3d0001 */
- 1.041450142840, /* 0x1.0a9c79ffe */
- 1.044273972530, /* 0x1.0b558a001 */
- 1.047105550795, /* 0x1.0c0f1c001 */
- 1.049944162390, /* 0x1.0cc924001 */
- 1.052791833895, /* 0x1.0d83c4001 */
- 1.055645227426, /* 0x1.0e3ec3fff */
- 1.058507919326, /* 0x1.0efa60001 */
- 1.061377286898, /* 0x1.0fb66bfff */
- 1.064254641510, /* 0x1.1072fdffe */
- 1.067140102389, /* 0x1.113018000 */
- 1.070034146304, /* 0x1.11edc1fff */
- 1.072937250162, /* 0x1.12ac04001 */
- 1.075843691823, /* 0x1.136a7dfff */
- 1.078760385496, /* 0x1.1429a3ffe */
- 1.081685543070, /* 0x1.14e958000 */
- 1.084618330005, /* 0x1.15a98c000 */
- 1.087556362176, /* 0x1.166a18001 */
- 1.090508937863, /* 0x1.172b98001 */
- 1.093464612954, /* 0x1.17ed4bfff */
- 1.096430182434, /* 0x1.18afa5ffe */
- 1.099401354802, /* 0x1.19725e000 */
- 1.102381587017, /* 0x1.1a35adfff */
- 1.105370759965, /* 0x1.1af994000 */
- 1.108367800686, /* 0x1.1bbdfdffe */
- 1.111373305331, /* 0x1.1c82f6000 */
- 1.114387035385, /* 0x1.1d4878001 */
- 1.117408752440, /* 0x1.1e0e7ffff */
- 1.120437502874, /* 0x1.1ed4fe000 */
- 1.123474478729, /* 0x1.1f9c06000 */
- 1.126521706601, /* 0x1.2063ba001 */
- 1.129574775716, /* 0x1.212bd0001 */
- 1.132638812065, /* 0x1.21f49e000 */
- 1.135709524130, /* 0x1.22bddbffe */
- 1.138789534565, /* 0x1.2387b5fff */
- 1.141876101508, /* 0x1.2451fe000 */
- 1.144971728301, /* 0x1.251cddffe */
- 1.148077130296, /* 0x1.25e861ffe */
- 1.151189923305, /* 0x1.26b462001 */
- 1.154312610610, /* 0x1.278107ffe */
- 1.157440662410, /* 0x1.284e08001 */
- 1.160578370109, /* 0x1.291baa001 */
- 1.163725256932, /* 0x1.29e9e6000 */
- 1.166879892324, /* 0x1.2ab8a3ffe */
- 1.170044302935, /* 0x1.2b8805fff */
- 1.173205971694, /* 0x1.2c5739ffe */
- 1.176397800428, /* 0x1.2d2867ffe */
- 1.179586529747, /* 0x1.2df962001 */
- 1.182784795737, /* 0x1.2ecafbffe */
- 1.185991406414, /* 0x1.2f9d21ffe */
- 1.189206838636, /* 0x1.306fdc001 */
- 1.192430973067, /* 0x1.314328000 */
- 1.195664167430, /* 0x1.32170c001 */
- 1.198906540890, /* 0x1.32eb8a001 */
- 1.202157497408, /* 0x1.33c098000 */
- 1.205416083326, /* 0x1.349625fff */
- 1.208683252332, /* 0x1.356c43fff */
- 1.211961269402, /* 0x1.364318001 */
- 1.215246438983, /* 0x1.371a64000 */
- 1.218539118740, /* 0x1.37f22dffe */
- 1.221847295770, /* 0x1.38cafc000 */
- 1.225158572187, /* 0x1.39a3fdfff */
- 1.228481650325, /* 0x1.3a7dc5ffe */
- 1.231811761846, /* 0x1.3b5803fff */
- 1.235149741144, /* 0x1.3c32c5ffe */
- 1.238499879811, /* 0x1.3d0e53ffe */
- 1.241858124726, /* 0x1.3dea69fff */
- 1.245225191102, /* 0x1.3ec713fff */
- 1.248601436624, /* 0x1.3fa458000 */
- 1.251975655584, /* 0x1.40817a001 */
- 1.255380749731, /* 0x1.4160a2001 */
- 1.258783102010, /* 0x1.423f9bffe */
- 1.262198328973, /* 0x1.431f6e000 */
- 1.265619754780, /* 0x1.43ffa7fff */
- 1.269052743928, /* 0x1.44e0a4001 */
- 1.272490739830, /* 0x1.45c1f4000 */
- 1.275942921659, /* 0x1.46a432001 */
- 1.279397487615, /* 0x1.478697ffe */
- 1.282870173427, /* 0x1.486a2dffe */
- 1.286346316319, /* 0x1.494dfdffe */
- 1.289836049094, /* 0x1.4a32b2001 */
- 1.293333172770, /* 0x1.4b17e1ffe */
- 1.296839594835, /* 0x1.4bfdadfff */
- 1.300354957560, /* 0x1.4ce40fffe */
- 1.303882122055, /* 0x1.4dcb38001 */
- 1.307417988757, /* 0x1.4eb2f1ffe */
- 1.310960650439, /* 0x1.4f9b1dfff */
- 1.314516782746, /* 0x1.50842bfff */
- 1.318079948424, /* 0x1.516daffff */
- 1.321653246888, /* 0x1.5257de000 */
- 1.325237751030, /* 0x1.5342c8001 */
- 1.328829526907, /* 0x1.542e2c000 */
- 1.332433700535, /* 0x1.551a5fffe */
- 1.336045145966, /* 0x1.56070dffe */
- 1.339667558645, /* 0x1.56f473ffe */
- 1.343300342533, /* 0x1.57e287ffe */
- 1.346941947961, /* 0x1.58d130001 */
- 1.350594043714, /* 0x1.59c087ffe */
- 1.354256033883, /* 0x1.5ab085fff */
- 1.357932448365, /* 0x1.5ba175ffe */
- 1.361609339707, /* 0x1.5c926dfff */
- 1.365299344044, /* 0x1.5d8441ffe */
- 1.369003057507, /* 0x1.5e76fc001 */
- 1.372714757920, /* 0x1.5f6a3c000 */
- 1.376437187179, /* 0x1.605e2fffe */
- 1.380165219333, /* 0x1.615282001 */
- 1.383909463864, /* 0x1.6247e3ffe */
- 1.387661933907, /* 0x1.633dd0000 */
- 1.391424179060, /* 0x1.64345fffe */
- 1.395197510706, /* 0x1.652ba9fff */
- 1.399006724329, /* 0x1.66254dffe */
- 1.402773022651, /* 0x1.671c22000 */
- 1.406576037403, /* 0x1.68155dfff */
- 1.410389423392, /* 0x1.690f48001 */
-};
-- 
2.11.0


^ permalink raw reply	[flat|nested] 19+ messages in thread

* [PATCH 2/4] Move exp compat wrappers under math/
  2017-09-13 10:43 [PATCH 0/4] Optimized generic expf and exp2f Szabolcs Nagy
  2017-09-13 10:45 ` [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers Szabolcs Nagy
@ 2017-09-13 10:46 ` Szabolcs Nagy
  2017-09-13 10:49   ` Szabolcs Nagy
  2017-09-13 10:47 ` [PATCH 3/4] New expf and exp2f version without SVID compat wrapper Szabolcs Nagy
                   ` (2 subsequent siblings)
  4 siblings, 1 reply; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 10:46 UTC (permalink / raw)
  To: GNU C Library; +Cc: nd

Move exp compat wrappers to math/w_exp{,f,l}_compat.c to be
consistent with other wrappers.

2017-09-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

	* sysdeps/ieee754/dbl-64/w_exp_compat.c: Move to...
	* math/w_exp_compat.c: ... here.
	* sysdeps/ieee754/flt-32/w_expf_compat.c: Move to...
	* math/w_expf_compat.c: ... here.
	* sysdeps/ieee754/ldbl-128/w_expl_compat.c: Move to...
	* math/w_expl_compat.c: ... here.
	* sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c: Remove.
	* sysdeps/ieee754/ldbl-96/w_expl_compat.c: Remove.
	* sysdeps/ieee754/ldbl-opt/w_exp_compat.c: Use the new path.
	* sysdeps/ieee754/ldbl-opt/w_expl_compat.c: Likewise.

^ permalink raw reply	[flat|nested] 19+ messages in thread

* [PATCH 3/4] New expf and exp2f version without SVID compat wrapper
  2017-09-13 10:43 [PATCH 0/4] Optimized generic expf and exp2f Szabolcs Nagy
  2017-09-13 10:45 ` [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers Szabolcs Nagy
  2017-09-13 10:46 ` [PATCH 2/4] Move exp compat wrappers under math/ Szabolcs Nagy
@ 2017-09-13 10:47 ` Szabolcs Nagy
  2017-09-13 20:17   ` Joseph Myers
  2017-09-13 10:48 ` [PATCH 4/4] Do not wrap expf and exp2f Szabolcs Nagy
  2017-09-13 13:43 ` [PATCH 0/4] Optimized generic " H.J. Lu
  4 siblings, 1 reply; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 10:47 UTC (permalink / raw)
  To: GNU C Library; +Cc: nd

[-- Attachment #1: Type: text/plain, Size: 2465 bytes --]

This patch changes the expf and exp2f error handling
semantics to only set errno accoring to POSIX rules.
New symbol version is introduced at GLIBC_2.27.

The old wrappers are kept for compat symbols (with
GLIBC_2_0 version and unchanged error semantics).

Internal calls to __expf now get the new error semantics,
this seems to only affect sysdeps/i386/fpu/s_expm1f.S
where the errno-only behaviour should be correct.

All linux libm abilists are updated (even ones where i
could not build a working toolchain).

2017-09-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

	* math/Versions (expf): New libm symbol at GLIBC_2.27.
	(exp2f): Likewise.
	* math/w_exp2f.c: New file.
	* math/w_expf.c: New file.
	* math/w_exp2f_compat.c (__exp2f_compat): For compat symbol only.
	* math/w_expf_compat.c (__expf_compat): Likewise.
	* sysdeps/unix/sysv/linux/aarch64/libm.abilist: Update.
	* sysdeps/unix/sysv/linux/alpha/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/arm/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/hppa/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/i386/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/ia64/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/microblaze/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/mips/mips32/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/mips/mips64/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/nios2/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist:
	Likewise.
	* sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist:
	Likewise.
	* sysdeps/unix/sysv/linux/powerpc/powerpc64/libm-le.abilist:
	Likewise.
	* sysdeps/unix/sysv/linux/powerpc/powerpc64/libm.abilist:
	Likewise.
	* sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/sh/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/tile/tilegx/tilegx32/libm.abilist:
	Likewise.
	* sysdeps/unix/sysv/linux/tile/tilegx/tilegx64/libm.abilist:
	Likewise.
	* sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/x86_64/64/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist: Likewise.

[-- Attachment #2: 0003-errno-only.patch --]
[-- Type: text/x-patch, Size: 17131 bytes --]

From f108f9ef952e9b46f89da762794b7220d34fa927 Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <szabolcs.nagy@arm.com>
Date: Tue, 12 Sep 2017 12:20:50 +0100
Subject: [PATCH 3/4] errno-only

---
 math/Versions                                                | 3 +++
 math/w_exp2f.c                                               | 7 +++++++
 math/w_exp2f_compat.c                                        | 6 +++---
 math/w_expf.c                                                | 7 +++++++
 math/w_expf_compat.c                                         | 7 +++----
 sysdeps/unix/sysv/linux/aarch64/libm.abilist                 | 3 +++
 sysdeps/unix/sysv/linux/alpha/libm.abilist                   | 3 +++
 sysdeps/unix/sysv/linux/arm/libm.abilist                     | 3 +++
 sysdeps/unix/sysv/linux/hppa/libm.abilist                    | 3 +++
 sysdeps/unix/sysv/linux/i386/libm.abilist                    | 3 +++
 sysdeps/unix/sysv/linux/ia64/libm.abilist                    | 3 +++
 sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist           | 3 +++
 sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist             | 3 +++
 sysdeps/unix/sysv/linux/microblaze/libm.abilist              | 3 +++
 sysdeps/unix/sysv/linux/mips/mips32/libm.abilist             | 3 +++
 sysdeps/unix/sysv/linux/mips/mips64/libm.abilist             | 3 +++
 sysdeps/unix/sysv/linux/nios2/libm.abilist                   | 3 +++
 sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist   | 3 +++
 sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist | 3 +++
 sysdeps/unix/sysv/linux/powerpc/powerpc64/libm-le.abilist    | 3 +++
 sysdeps/unix/sysv/linux/powerpc/powerpc64/libm.abilist       | 3 +++
 sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist            | 3 +++
 sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist            | 3 +++
 sysdeps/unix/sysv/linux/sh/libm.abilist                      | 3 +++
 sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist           | 3 +++
 sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist           | 3 +++
 sysdeps/unix/sysv/linux/tile/tilegx/tilegx32/libm.abilist    | 3 +++
 sysdeps/unix/sysv/linux/tile/tilegx/tilegx64/libm.abilist    | 3 +++
 sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist            | 3 +++
 sysdeps/unix/sysv/linux/x86_64/64/libm.abilist               | 3 +++
 sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist              | 3 +++
 31 files changed, 101 insertions(+), 7 deletions(-)
 create mode 100644 math/w_exp2f.c
 create mode 100644 math/w_expf.c

diff --git a/math/Versions b/math/Versions
index e409de0681..380f6a2a1a 100644
--- a/math/Versions
+++ b/math/Versions
@@ -229,4 +229,7 @@ libm {
     fromfp; fromfpf; fromfpl; ufromfp; ufromfpf; ufromfpl;
     fromfpx; fromfpxf; fromfpxl; ufromfpx; ufromfpxf; ufromfpxl;
   }
+  GLIBC_2.27 {
+    expf; exp2f;
+  }
 }
diff --git a/math/w_exp2f.c b/math/w_exp2f.c
new file mode 100644
index 0000000000..948d745dab
--- /dev/null
+++ b/math/w_exp2f.c
@@ -0,0 +1,7 @@
+#include <math-type-macros-float.h>
+#undef __USE_WRAPPER_TEMPLATE
+#define __USE_WRAPPER_TEMPLATE 1
+#undef declare_mgen_alias
+#define declare_mgen_alias(a, b)
+#include <w_exp2_template.c>
+versioned_symbol (libm, __exp2f, exp2f, GLIBC_2_27);
diff --git a/math/w_exp2f_compat.c b/math/w_exp2f_compat.c
index b058daed30..6ab6f6c6de 100644
--- a/math/w_exp2f_compat.c
+++ b/math/w_exp2f_compat.c
@@ -6,9 +6,9 @@
 #include <math_private.h>
 #include <math-svid-compat.h>
 
-#if LIBM_SVID_COMPAT
+#if LIBM_SVID_COMPAT && SHLIB_COMPAT (libm, GLIBC_2_1, GLIBC_2_27)
 float
-__exp2f (float x)
+__exp2f_compat (float x)
 {
   float z = __ieee754_exp2f (x);
   if (__builtin_expect (!isfinite (z) || z == 0, 0)
@@ -18,5 +18,5 @@ __exp2f (float x)
 
   return z;
 }
-weak_alias (__exp2f, exp2f)
+compat_symbol (libm, __exp2f_compat, exp2f, GLIBC_2_1);
 #endif
diff --git a/math/w_expf.c b/math/w_expf.c
new file mode 100644
index 0000000000..6d5f03c5f9
--- /dev/null
+++ b/math/w_expf.c
@@ -0,0 +1,7 @@
+#include <math-type-macros-float.h>
+#undef __USE_WRAPPER_TEMPLATE
+#define __USE_WRAPPER_TEMPLATE 1
+#undef declare_mgen_alias
+#define declare_mgen_alias(a, b)
+#include <w_exp_template.c>
+versioned_symbol (libm, __expf, expf, GLIBC_2_27);
diff --git a/math/w_expf_compat.c b/math/w_expf_compat.c
index a38ff4082d..2512498bb3 100644
--- a/math/w_expf_compat.c
+++ b/math/w_expf_compat.c
@@ -20,10 +20,10 @@
 #include <math_private.h>
 #include <math-svid-compat.h>
 
-#if LIBM_SVID_COMPAT
+#if LIBM_SVID_COMPAT && SHLIB_COMPAT (libm, GLIBC_2_0, GLIBC_2_27)
 /* wrapper expf */
 float
-__expf (float x)
+__expf_compat (float x)
 {
   float z = __ieee754_expf (x);
   if (__builtin_expect (!isfinite (z) || z == 0, 0)
@@ -32,6 +32,5 @@ __expf (float x)
 
   return z;
 }
-hidden_def (__expf)
-weak_alias (__expf, expf)
+compat_symbol (libm, __expf_compat, expf, GLIBC_2_0);
 #endif
diff --git a/sysdeps/unix/sysv/linux/aarch64/libm.abilist b/sysdeps/unix/sysv/linux/aarch64/libm.abilist
index 6266864ea7..10102eeaff 100644
--- a/sysdeps/unix/sysv/linux/aarch64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/aarch64/libm.abilist
@@ -460,3 +460,6 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/alpha/libm.abilist b/sysdeps/unix/sysv/linux/alpha/libm.abilist
index 0dc44afd77..e09a115aa9 100644
--- a/sysdeps/unix/sysv/linux/alpha/libm.abilist
+++ b/sysdeps/unix/sysv/linux/alpha/libm.abilist
@@ -470,6 +470,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.3.4 GLIBC_2.3.4 A
 GLIBC_2.3.4 __c1_cabsf F
 GLIBC_2.3.4 __c1_cacosf F
diff --git a/sysdeps/unix/sysv/linux/arm/libm.abilist b/sysdeps/unix/sysv/linux/arm/libm.abilist
index d0c33d2a11..8095876449 100644
--- a/sysdeps/unix/sysv/linux/arm/libm.abilist
+++ b/sysdeps/unix/sysv/linux/arm/libm.abilist
@@ -117,6 +117,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 _LIB_VERSION D 0x4
 GLIBC_2.4 __clog10 F
diff --git a/sysdeps/unix/sysv/linux/hppa/libm.abilist b/sysdeps/unix/sysv/linux/hppa/libm.abilist
index 80fd4ada99..19d40ef50d 100644
--- a/sysdeps/unix/sysv/linux/hppa/libm.abilist
+++ b/sysdeps/unix/sysv/linux/hppa/libm.abilist
@@ -429,5 +429,8 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 exp2l F
diff --git a/sysdeps/unix/sysv/linux/i386/libm.abilist b/sysdeps/unix/sysv/linux/i386/libm.abilist
index 2aef7e5aed..791fba28e5 100644
--- a/sysdeps/unix/sysv/linux/i386/libm.abilist
+++ b/sysdeps/unix/sysv/linux/i386/libm.abilist
@@ -611,4 +611,7 @@ GLIBC_2.26 ufromfpxf128 F
 GLIBC_2.26 y0f128 F
 GLIBC_2.26 y1f128 F
 GLIBC_2.26 ynf128 F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
diff --git a/sysdeps/unix/sysv/linux/ia64/libm.abilist b/sysdeps/unix/sysv/linux/ia64/libm.abilist
index f165eb3957..65a0fbe56a 100644
--- a/sysdeps/unix/sysv/linux/ia64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/ia64/libm.abilist
@@ -540,4 +540,7 @@ GLIBC_2.26 ufromfpxf128 F
 GLIBC_2.26 y0f128 F
 GLIBC_2.26 y1f128 F
 GLIBC_2.26 ynf128 F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
diff --git a/sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist b/sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist
index d0c33d2a11..8095876449 100644
--- a/sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist
+++ b/sysdeps/unix/sysv/linux/m68k/coldfire/libm.abilist
@@ -117,6 +117,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 _LIB_VERSION D 0x4
 GLIBC_2.4 __clog10 F
diff --git a/sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist b/sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist
index f3de018cd9..5e692dda7b 100644
--- a/sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist
+++ b/sysdeps/unix/sysv/linux/m68k/m680x0/libm.abilist
@@ -471,4 +471,7 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
diff --git a/sysdeps/unix/sysv/linux/microblaze/libm.abilist b/sysdeps/unix/sysv/linux/microblaze/libm.abilist
index bb872f292a..65f1d5b451 100644
--- a/sysdeps/unix/sysv/linux/microblaze/libm.abilist
+++ b/sysdeps/unix/sysv/linux/microblaze/libm.abilist
@@ -428,3 +428,6 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/mips/mips32/libm.abilist b/sysdeps/unix/sysv/linux/mips/mips32/libm.abilist
index 000c011495..c32ea5b96a 100644
--- a/sysdeps/unix/sysv/linux/mips/mips32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/mips/mips32/libm.abilist
@@ -430,6 +430,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 exp2l F
 _gp_disp _gp_disp A
diff --git a/sysdeps/unix/sysv/linux/mips/mips64/libm.abilist b/sysdeps/unix/sysv/linux/mips/mips64/libm.abilist
index b9599fee34..18b2aa2404 100644
--- a/sysdeps/unix/sysv/linux/mips/mips64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/mips/mips64/libm.abilist
@@ -462,4 +462,7 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
diff --git a/sysdeps/unix/sysv/linux/nios2/libm.abilist b/sysdeps/unix/sysv/linux/nios2/libm.abilist
index c858ed5907..e492a68e9d 100644
--- a/sysdeps/unix/sysv/linux/nios2/libm.abilist
+++ b/sysdeps/unix/sysv/linux/nios2/libm.abilist
@@ -428,3 +428,6 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist b/sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist
index 13ea493ac8..ad8f0372f4 100644
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libm.abilist
@@ -473,6 +473,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 __clog10l F
 GLIBC_2.4 __finitel F
diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist b/sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist
index ff6dc7a361..9c26b5b809 100644
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libm.abilist
@@ -472,6 +472,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 __clog10l F
 GLIBC_2.4 __finitel F
diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm-le.abilist b/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm-le.abilist
index d79cb99325..8e36699f28 100644
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm-le.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm-le.abilist
@@ -605,3 +605,6 @@ GLIBC_2.26 ufromfpxf128 F
 GLIBC_2.26 y0f128 F
 GLIBC_2.26 y1f128 F
 GLIBC_2.26 ynf128 F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm.abilist b/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm.abilist
index afd381791a..9ca0c3ccfc 100644
--- a/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc64/libm.abilist
@@ -148,6 +148,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.3 GLIBC_2.3 A
 GLIBC_2.3 _LIB_VERSION D 0x4
 GLIBC_2.3 __clog10 F
diff --git a/sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist b/sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist
index 604dff18d4..8a79f0137f 100644
--- a/sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/s390/s390-32/libm.abilist
@@ -460,6 +460,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 __clog10l F
 GLIBC_2.4 __finitel F
diff --git a/sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist b/sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist
index b56701e350..df81853618 100644
--- a/sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/s390/s390-64/libm.abilist
@@ -458,6 +458,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 __clog10l F
 GLIBC_2.4 __finitel F
diff --git a/sysdeps/unix/sysv/linux/sh/libm.abilist b/sysdeps/unix/sysv/linux/sh/libm.abilist
index 6b24852b68..6b6a42dc9c 100644
--- a/sysdeps/unix/sysv/linux/sh/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sh/libm.abilist
@@ -429,5 +429,8 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 exp2l F
diff --git a/sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist b/sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist
index 61042ba147..24d67d22e1 100644
--- a/sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sparc/sparc32/libm.abilist
@@ -464,6 +464,9 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
 GLIBC_2.4 __clog10l F
 GLIBC_2.4 __finitel F
diff --git a/sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist b/sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist
index d071835880..2fdccc0de3 100644
--- a/sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/sparc/sparc64/libm.abilist
@@ -461,4 +461,7 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
diff --git a/sysdeps/unix/sysv/linux/tile/tilegx/tilegx32/libm.abilist b/sysdeps/unix/sysv/linux/tile/tilegx/tilegx32/libm.abilist
index 0059c27fa2..98bc348f91 100644
--- a/sysdeps/unix/sysv/linux/tile/tilegx/tilegx32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/tile/tilegx/tilegx32/libm.abilist
@@ -429,3 +429,6 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/tile/tilegx/tilegx64/libm.abilist b/sysdeps/unix/sysv/linux/tile/tilegx/tilegx64/libm.abilist
index 0059c27fa2..98bc348f91 100644
--- a/sysdeps/unix/sysv/linux/tile/tilegx/tilegx64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/tile/tilegx/tilegx64/libm.abilist
@@ -429,3 +429,6 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist b/sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist
index 0059c27fa2..98bc348f91 100644
--- a/sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist
+++ b/sysdeps/unix/sysv/linux/tile/tilepro/libm.abilist
@@ -429,3 +429,6 @@ GLIBC_2.25 ufromfpl F
 GLIBC_2.25 ufromfpx F
 GLIBC_2.25 ufromfpxf F
 GLIBC_2.25 ufromfpxl F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
diff --git a/sysdeps/unix/sysv/linux/x86_64/64/libm.abilist b/sysdeps/unix/sysv/linux/x86_64/64/libm.abilist
index 44473c2cbe..e6fd3fe3df 100644
--- a/sysdeps/unix/sysv/linux/x86_64/64/libm.abilist
+++ b/sysdeps/unix/sysv/linux/x86_64/64/libm.abilist
@@ -600,4 +600,7 @@ GLIBC_2.26 ufromfpxf128 F
 GLIBC_2.26 y0f128 F
 GLIBC_2.26 y1f128 F
 GLIBC_2.26 ynf128 F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
 GLIBC_2.4 GLIBC_2.4 A
diff --git a/sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist b/sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist
index cc19a73112..afa7b98697 100644
--- a/sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist
+++ b/sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist
@@ -599,3 +599,6 @@ GLIBC_2.26 ufromfpxf128 F
 GLIBC_2.26 y0f128 F
 GLIBC_2.26 y1f128 F
 GLIBC_2.26 ynf128 F
+GLIBC_2.27 GLIBC_2.27 A
+GLIBC_2.27 exp2f F
+GLIBC_2.27 expf F
-- 
2.11.0


^ permalink raw reply	[flat|nested] 19+ messages in thread

* [PATCH 4/4] Do not wrap expf and exp2f
  2017-09-13 10:43 [PATCH 0/4] Optimized generic expf and exp2f Szabolcs Nagy
                   ` (2 preceding siblings ...)
  2017-09-13 10:47 ` [PATCH 3/4] New expf and exp2f version without SVID compat wrapper Szabolcs Nagy
@ 2017-09-13 10:48 ` Szabolcs Nagy
  2017-09-13 20:19   ` Joseph Myers
  2017-09-13 13:43 ` [PATCH 0/4] Optimized generic " H.J. Lu
  4 siblings, 1 reply; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 10:48 UTC (permalink / raw)
  To: GNU C Library; +Cc: nd

[-- Attachment #1: Type: text/plain, Size: 1022 bytes --]

The new generic expf and exp2f code don't need wrappers any
more, they set errno inline, so only use the wrappers on
targets that need it. (The top level wrapper code is included
into sysdeps dirs where the wrapper is needed and empty
w_exp*f.c where the wrapper is suppressed).

A powerpc64 expf implementation includes the expf c code
directly which needed some changes.

2017-09-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>

	* sysdeps/ieee754/flt-32/e_exp2f.c (__exp2f): Define without wrapper.
	* sysdeps/ieee754/flt-32/e_expf.c (__expf): Likewise
	* sysdeps/ieee754/flt-32/w_exp2f.c: New file.
	* sysdeps/ieee754/flt-32/w_expf.c: New file.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c: Update for
	the new expf code.
	* sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c: New file.
	* sysdeps/m68k/m680x0/fpu/w_expf.c: New file.
	* sysdeps/i386/fpu/w_exp2f.c: New file.
	* sysdeps/i386/fpu/w_expf.c: New file.
	* sysdeps/i386/i686/fpu/multiarch/w_expf.c: New file.
	* sysdeps/x86_64/fpu/w_expf.c: New file.

[-- Attachment #2: 0004-no-wrapper.patch --]
[-- Type: text/x-patch, Size: 5323 bytes --]

From ba8143da8b2d0a25e1c4ed42704d46a720ebc3e1 Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <szabolcs.nagy@arm.com>
Date: Tue, 12 Sep 2017 12:44:18 +0100
Subject: [PATCH 4/4] no wrapper

---
 sysdeps/i386/fpu/w_exp2f.c                             | 1 +
 sysdeps/i386/fpu/w_expf.c                              | 1 +
 sysdeps/i386/i686/fpu/multiarch/w_expf.c               | 1 +
 sysdeps/ieee754/flt-32/e_exp2f.c                       | 7 +++++--
 sysdeps/ieee754/flt-32/e_expf.c                        | 8 ++++++--
 sysdeps/ieee754/flt-32/w_exp2f.c                       | 1 +
 sysdeps/ieee754/flt-32/w_expf.c                        | 1 +
 sysdeps/m68k/m680x0/fpu/w_expf.c                       | 1 +
 sysdeps/powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c | 7 ++++++-
 sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c       | 1 +
 sysdeps/x86_64/fpu/w_expf.c                            | 1 +
 11 files changed, 25 insertions(+), 5 deletions(-)
 create mode 100644 sysdeps/i386/fpu/w_exp2f.c
 create mode 100644 sysdeps/i386/fpu/w_expf.c
 create mode 100644 sysdeps/i386/i686/fpu/multiarch/w_expf.c
 create mode 100644 sysdeps/ieee754/flt-32/w_exp2f.c
 create mode 100644 sysdeps/ieee754/flt-32/w_expf.c
 create mode 100644 sysdeps/m68k/m680x0/fpu/w_expf.c
 create mode 100644 sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c
 create mode 100644 sysdeps/x86_64/fpu/w_expf.c

diff --git a/sysdeps/i386/fpu/w_exp2f.c b/sysdeps/i386/fpu/w_exp2f.c
new file mode 100644
index 0000000000..4418f8bed5
--- /dev/null
+++ b/sysdeps/i386/fpu/w_exp2f.c
@@ -0,0 +1 @@
+#include <math/w_exp2f.c>
diff --git a/sysdeps/i386/fpu/w_expf.c b/sysdeps/i386/fpu/w_expf.c
new file mode 100644
index 0000000000..b471bccdb1
--- /dev/null
+++ b/sysdeps/i386/fpu/w_expf.c
@@ -0,0 +1 @@
+#include <math/w_expf.c>
diff --git a/sysdeps/i386/i686/fpu/multiarch/w_expf.c b/sysdeps/i386/i686/fpu/multiarch/w_expf.c
new file mode 100644
index 0000000000..b471bccdb1
--- /dev/null
+++ b/sysdeps/i386/i686/fpu/multiarch/w_expf.c
@@ -0,0 +1 @@
+#include <math/w_expf.c>
diff --git a/sysdeps/ieee754/flt-32/e_exp2f.c b/sysdeps/ieee754/flt-32/e_exp2f.c
index be53e5c699..7a3c5fdcff 100644
--- a/sysdeps/ieee754/flt-32/e_exp2f.c
+++ b/sysdeps/ieee754/flt-32/e_exp2f.c
@@ -18,6 +18,7 @@
 
 #include <math.h>
 #include <stdint.h>
+#include <shlib-compat.h>
 #include "math_config.h"
 
 /*
@@ -42,7 +43,7 @@ top12 (float x)
 }
 
 float
-__ieee754_exp2f (float x)
+__exp2f (float x)
 {
   uint32_t abstop;
   uint64_t ki, t;
@@ -85,4 +86,6 @@ __ieee754_exp2f (float x)
   y = y * s;
   return (float) y;
 }
-strong_alias (__ieee754_exp2f, __exp2f_finite)
+strong_alias (__exp2f, __ieee754_exp2f)
+strong_alias (__exp2f, __exp2f_finite)
+versioned_symbol (libm, __exp2f, exp2f, GLIBC_2_27);
diff --git a/sysdeps/ieee754/flt-32/e_expf.c b/sysdeps/ieee754/flt-32/e_expf.c
index b8244a82cf..8a38c2a0c6 100644
--- a/sysdeps/ieee754/flt-32/e_expf.c
+++ b/sysdeps/ieee754/flt-32/e_expf.c
@@ -18,6 +18,7 @@
 
 #include <math.h>
 #include <stdint.h>
+#include <shlib-compat.h>
 #include "math_config.h"
 
 /*
@@ -42,7 +43,7 @@ top12 (float x)
 }
 
 float
-__ieee754_expf (float x)
+__expf (float x)
 {
   uint32_t abstop;
   uint64_t ki, t;
@@ -99,4 +100,7 @@ __ieee754_expf (float x)
   y = y * s;
   return (float) y;
 }
-strong_alias (__ieee754_expf, __expf_finite)
+hidden_def (__expf)
+strong_alias (__expf, __ieee754_expf)
+strong_alias (__expf, __expf_finite)
+versioned_symbol (libm, __expf, expf, GLIBC_2_27);
diff --git a/sysdeps/ieee754/flt-32/w_exp2f.c b/sysdeps/ieee754/flt-32/w_exp2f.c
new file mode 100644
index 0000000000..1cc8931700
--- /dev/null
+++ b/sysdeps/ieee754/flt-32/w_exp2f.c
@@ -0,0 +1 @@
+/* Not needed.  */
diff --git a/sysdeps/ieee754/flt-32/w_expf.c b/sysdeps/ieee754/flt-32/w_expf.c
new file mode 100644
index 0000000000..1cc8931700
--- /dev/null
+++ b/sysdeps/ieee754/flt-32/w_expf.c
@@ -0,0 +1 @@
+/* Not needed.  */
diff --git a/sysdeps/m68k/m680x0/fpu/w_expf.c b/sysdeps/m68k/m680x0/fpu/w_expf.c
new file mode 100644
index 0000000000..b471bccdb1
--- /dev/null
+++ b/sysdeps/m68k/m680x0/fpu/w_expf.c
@@ -0,0 +1 @@
+#include <math/w_expf.c>
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c b/sysdeps/powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c
index b236290ea2..cc23f06fb1 100644
--- a/sysdeps/powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c
+++ b/sysdeps/powerpc/powerpc64/fpu/multiarch/e_expf-ppc64.c
@@ -16,9 +16,14 @@
    License along with the GNU C Library; if not, see
    <http://www.gnu.org/licenses/>.  */
 
+#include <shlib-compat.h>
 #undef strong_alias
 #define strong_alias(a, b)
+#undef hidden_def
+#define hidden_def(a)
+#undef versioned_symbol
+#define versioned_symbol(a, b, c, d)
 
-#define __ieee754_expf __ieee754_expf_ppc64
+#define __expf __ieee754_expf_ppc64
 
 #include <sysdeps/ieee754/flt-32/e_expf.c>
diff --git a/sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c b/sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c
new file mode 100644
index 0000000000..b471bccdb1
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/fpu/multiarch/w_expf.c
@@ -0,0 +1 @@
+#include <math/w_expf.c>
diff --git a/sysdeps/x86_64/fpu/w_expf.c b/sysdeps/x86_64/fpu/w_expf.c
new file mode 100644
index 0000000000..b471bccdb1
--- /dev/null
+++ b/sysdeps/x86_64/fpu/w_expf.c
@@ -0,0 +1 @@
+#include <math/w_expf.c>
-- 
2.11.0


^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 2/4] Move exp compat wrappers under math/
  2017-09-13 10:46 ` [PATCH 2/4] Move exp compat wrappers under math/ Szabolcs Nagy
@ 2017-09-13 10:49   ` Szabolcs Nagy
  2017-09-13 12:43     ` Joseph Myers
  0 siblings, 1 reply; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 10:49 UTC (permalink / raw)
  To: GNU C Library; +Cc: nd

[-- Attachment #1: Type: text/plain, Size: 737 bytes --]

On 13/09/17 11:46, Szabolcs Nagy wrote:
> Move exp compat wrappers to math/w_exp{,f,l}_compat.c to be
> consistent with other wrappers.
> 
> 2017-09-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>
> 
> 	* sysdeps/ieee754/dbl-64/w_exp_compat.c: Move to...
> 	* math/w_exp_compat.c: ... here.
> 	* sysdeps/ieee754/flt-32/w_expf_compat.c: Move to...
> 	* math/w_expf_compat.c: ... here.
> 	* sysdeps/ieee754/ldbl-128/w_expl_compat.c: Move to...
> 	* math/w_expl_compat.c: ... here.
> 	* sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c: Remove.
> 	* sysdeps/ieee754/ldbl-96/w_expl_compat.c: Remove.
> 	* sysdeps/ieee754/ldbl-opt/w_exp_compat.c: Use the new path.
> 	* sysdeps/ieee754/ldbl-opt/w_expl_compat.c: Likewise.
> 

now with patch attached.


[-- Attachment #2: 0002-w_exp-_compat.patch --]
[-- Type: text/x-patch, Size: 7776 bytes --]

From a6ff6f68f2b610f6e78ba9a40c08a526f7b2cc7c Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <szabolcs.nagy@arm.com>
Date: Wed, 6 Sep 2017 16:48:41 +0100
Subject: [PATCH 2/4] w_exp*_compat

---
 {sysdeps/ieee754/dbl-64 => math}/w_exp_compat.c    |  0
 {sysdeps/ieee754/flt-32 => math}/w_expf_compat.c   |  0
 math/w_expl_compat.c                               | 48 ++++++++++++++++++----
 sysdeps/ieee754/ldbl-128/w_expl_compat.c           | 45 --------------------
 sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c        | 24 -----------
 sysdeps/ieee754/ldbl-96/w_expl_compat.c            | 37 -----------------
 sysdeps/ieee754/ldbl-opt/w_exp_compat.c            |  2 +-
 .../{ldbl-64-128 => ldbl-opt}/w_expl_compat.c      |  2 +-
 8 files changed, 42 insertions(+), 116 deletions(-)
 rename {sysdeps/ieee754/dbl-64 => math}/w_exp_compat.c (100%)
 rename {sysdeps/ieee754/flt-32 => math}/w_expf_compat.c (100%)
 delete mode 100644 sysdeps/ieee754/ldbl-128/w_expl_compat.c
 delete mode 100644 sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c
 delete mode 100644 sysdeps/ieee754/ldbl-96/w_expl_compat.c
 rename sysdeps/ieee754/{ldbl-64-128 => ldbl-opt}/w_expl_compat.c (72%)

diff --git a/sysdeps/ieee754/dbl-64/w_exp_compat.c b/math/w_exp_compat.c
similarity index 100%
rename from sysdeps/ieee754/dbl-64/w_exp_compat.c
rename to math/w_exp_compat.c
diff --git a/sysdeps/ieee754/flt-32/w_expf_compat.c b/math/w_expf_compat.c
similarity index 100%
rename from sysdeps/ieee754/flt-32/w_expf_compat.c
rename to math/w_expf_compat.c
diff --git a/math/w_expl_compat.c b/math/w_expl_compat.c
index 70096a820c..ec076deaad 100644
--- a/math/w_expl_compat.c
+++ b/math/w_expl_compat.c
@@ -1,13 +1,45 @@
+/* w_expl.c -- long double version of w_exp.c.
+ * Conversion to long double by Ulrich Drepper,
+ * Cygnus Support, drepper@cygnus.com.
+ */
+
+/*
+ * ====================================================
+ * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
+ *
+ * Developed at SunPro, a Sun Microsystems, Inc. business.
+ * Permission to use, copy, modify, and distribute this
+ * software is freely granted, provided that this notice
+ * is preserved.
+ * ====================================================
+ */
+
+#if defined(LIBM_SCCS) && !defined(lint)
+static char rcsid[] = "$NetBSD: $";
+#endif
+
+/*
+ * wrapper expl(x)
+ */
+
 #include <math.h>
-#include <stdio.h>
-#include <errno.h>
+#include <math_private.h>
+#include <math-svid-compat.h>
 
-long double
-__expl(long double x)
+#if LIBM_SVID_COMPAT
+long double __expl(long double x)	/* wrapper exp */
 {
-  fputs ("__expl not implemented\n", stderr);
-  __set_errno (ENOSYS);
-  return 0.0;
-}
+# ifdef _IEEE_LIBM
+	return __ieee754_expl(x);
+# else
+	long double z = __ieee754_expl (x);
+	if (__glibc_unlikely (!isfinite (z) || z == 0)
+	    && isfinite (x) && _LIB_VERSION != _IEEE_)
+	  return __kernel_standard_l (x, x, 206 + !!signbit (x));
 
+	return z;
+# endif
+}
+hidden_def (__expl)
 weak_alias (__expl, expl)
+#endif
diff --git a/sysdeps/ieee754/ldbl-128/w_expl_compat.c b/sysdeps/ieee754/ldbl-128/w_expl_compat.c
deleted file mode 100644
index ec076deaad..0000000000
--- a/sysdeps/ieee754/ldbl-128/w_expl_compat.c
+++ /dev/null
@@ -1,45 +0,0 @@
-/* w_expl.c -- long double version of w_exp.c.
- * Conversion to long double by Ulrich Drepper,
- * Cygnus Support, drepper@cygnus.com.
- */
-
-/*
- * ====================================================
- * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
- *
- * Developed at SunPro, a Sun Microsystems, Inc. business.
- * Permission to use, copy, modify, and distribute this
- * software is freely granted, provided that this notice
- * is preserved.
- * ====================================================
- */
-
-#if defined(LIBM_SCCS) && !defined(lint)
-static char rcsid[] = "$NetBSD: $";
-#endif
-
-/*
- * wrapper expl(x)
- */
-
-#include <math.h>
-#include <math_private.h>
-#include <math-svid-compat.h>
-
-#if LIBM_SVID_COMPAT
-long double __expl(long double x)	/* wrapper exp */
-{
-# ifdef _IEEE_LIBM
-	return __ieee754_expl(x);
-# else
-	long double z = __ieee754_expl (x);
-	if (__glibc_unlikely (!isfinite (z) || z == 0)
-	    && isfinite (x) && _LIB_VERSION != _IEEE_)
-	  return __kernel_standard_l (x, x, 206 + !!signbit (x));
-
-	return z;
-# endif
-}
-hidden_def (__expl)
-weak_alias (__expl, expl)
-#endif
diff --git a/sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c b/sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c
deleted file mode 100644
index 53948a96e6..0000000000
--- a/sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c
+++ /dev/null
@@ -1,24 +0,0 @@
-#include <math.h>
-#include <math_private.h>
-#include <math-svid-compat.h>
-#include <math_ldbl_opt.h>
-
-#if LIBM_SVID_COMPAT
-long double __expl(long double x)	/* wrapper exp  */
-{
-  long double z;
-  z = __ieee754_expl(x);
-  if (_LIB_VERSION == _IEEE_)
-    return z;
-  if (isfinite(x))
-    {
-      if (!isfinite (z))
-	return __kernel_standard_l(x,x,206); /* exp overflow  */
-      else if (z == 0.0L)
-	return __kernel_standard_l(x,x,207); /* exp underflow  */
-    }
-  return z;
-}
-hidden_def (__expl)
-long_double_symbol (libm, __expl, expl);
-#endif
diff --git a/sysdeps/ieee754/ldbl-96/w_expl_compat.c b/sysdeps/ieee754/ldbl-96/w_expl_compat.c
deleted file mode 100644
index ae9ab586cb..0000000000
--- a/sysdeps/ieee754/ldbl-96/w_expl_compat.c
+++ /dev/null
@@ -1,37 +0,0 @@
-/* Copyright (C) 2011-2017 Free Software Foundation, Inc.
-   This file is part of the GNU C Library.
-   Contributed by Ulrich Drepper <drepper@gmail.com>, 2011.
-
-   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
-   <http://www.gnu.org/licenses/>.  */
-
-#include <math.h>
-#include <math_private.h>
-#include <math-svid-compat.h>
-
-#if LIBM_SVID_COMPAT
-/* wrapper expl */
-long double
-__expl (long double x)
-{
-  long double z = __ieee754_expl (x);
-  if (__builtin_expect (!isfinite (z) || z == 0, 0)
-      && isfinite (x) && _LIB_VERSION != _IEEE_)
-    return __kernel_standard_l (x, x, 206 + !!signbit (x));
-
-  return z;
-}
-hidden_def (__expl)
-weak_alias (__expl, expl)
-#endif
diff --git a/sysdeps/ieee754/ldbl-opt/w_exp_compat.c b/sysdeps/ieee754/ldbl-opt/w_exp_compat.c
index 686c9c26d0..d01fe655ff 100644
--- a/sysdeps/ieee754/ldbl-opt/w_exp_compat.c
+++ b/sysdeps/ieee754/ldbl-opt/w_exp_compat.c
@@ -1,5 +1,5 @@
 #include <math_ldbl_opt.h>
-#include <sysdeps/ieee754/dbl-64/w_exp_compat.c>
+#include <math/w_exp_compat.c>
 #if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
 compat_symbol (libm, __exp, expl, GLIBC_2_0);
 #endif
diff --git a/sysdeps/ieee754/ldbl-64-128/w_expl_compat.c b/sysdeps/ieee754/ldbl-opt/w_expl_compat.c
similarity index 72%
rename from sysdeps/ieee754/ldbl-64-128/w_expl_compat.c
rename to sysdeps/ieee754/ldbl-opt/w_expl_compat.c
index 037e8bf348..4534051d88 100644
--- a/sysdeps/ieee754/ldbl-64-128/w_expl_compat.c
+++ b/sysdeps/ieee754/ldbl-opt/w_expl_compat.c
@@ -1,7 +1,7 @@
 #include <math_ldbl_opt.h>
 #undef weak_alias
 #define weak_alias(n,a)
-#include <sysdeps/ieee754/ldbl-128/w_expl_compat.c>
+#include <math/w_expl_compat.c>
 #if LIBM_SVID_COMPAT
 long_double_symbol (libm, __expl, expl);
 #endif
-- 
2.11.0


^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-13 10:45 ` [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers Szabolcs Nagy
@ 2017-09-13 12:41   ` Joseph Myers
  2017-09-13 20:13   ` Joseph Myers
  1 sibling, 0 replies; 19+ messages in thread
From: Joseph Myers @ 2017-09-13 12:41 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: GNU C Library, nd

On Wed, 13 Sep 2017, Szabolcs Nagy wrote:

> with the new expf benchmark
> https://sourceware.org/ml/libc-alpha/2017-08/msg01126.html

We need to get this reviewed.  Is it being pinged?  Are the benchtests 
maintainers looking at it?

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 2/4] Move exp compat wrappers under math/
  2017-09-13 10:49   ` Szabolcs Nagy
@ 2017-09-13 12:43     ` Joseph Myers
  0 siblings, 0 replies; 19+ messages in thread
From: Joseph Myers @ 2017-09-13 12:43 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: GNU C Library, nd

On Wed, 13 Sep 2017, Szabolcs Nagy wrote:

> On 13/09/17 11:46, Szabolcs Nagy wrote:
> > Move exp compat wrappers to math/w_exp{,f,l}_compat.c to be
> > consistent with other wrappers.
> > 
> > 2017-09-13  Szabolcs Nagy  <szabolcs.nagy@arm.com>
> > 
> > 	* sysdeps/ieee754/dbl-64/w_exp_compat.c: Move to...
> > 	* math/w_exp_compat.c: ... here.
> > 	* sysdeps/ieee754/flt-32/w_expf_compat.c: Move to...
> > 	* math/w_expf_compat.c: ... here.
> > 	* sysdeps/ieee754/ldbl-128/w_expl_compat.c: Move to...
> > 	* math/w_expl_compat.c: ... here.
> > 	* sysdeps/ieee754/ldbl-128ibm/w_expl_compat.c: Remove.
> > 	* sysdeps/ieee754/ldbl-96/w_expl_compat.c: Remove.
> > 	* sysdeps/ieee754/ldbl-opt/w_exp_compat.c: Use the new path.
> > 	* sysdeps/ieee754/ldbl-opt/w_expl_compat.c: Likewise.
> > 
> 
> now with patch attached.

OK.

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 0/4] Optimized generic expf and exp2f
  2017-09-13 10:43 [PATCH 0/4] Optimized generic expf and exp2f Szabolcs Nagy
                   ` (3 preceding siblings ...)
  2017-09-13 10:48 ` [PATCH 4/4] Do not wrap expf and exp2f Szabolcs Nagy
@ 2017-09-13 13:43 ` H.J. Lu
  2017-09-13 13:48   ` Szabolcs Nagy
  4 siblings, 1 reply; 19+ messages in thread
From: H.J. Lu @ 2017-09-13 13:43 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: GNU C Library, nd

On Wed, Sep 13, 2017 at 3:43 AM, Szabolcs Nagy <szabolcs.nagy@arm.com> wrote:
> broke down the expf patch into 4 parts.
> build-many-glibcs is still running i'll report the
> results once it's finished.
>
> Szabolcs Nagy (4):
>   Optimized generic expf and exp2f with wrappers
>   Move exp compat wrappers under math/
>   New expf and exp2f version without SVID compat wrapper
>   Do not wrap expf and exp2f

Do you have a glibc git branch I can try?

Thanks.

-- 
H.J.

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 0/4] Optimized generic expf and exp2f
  2017-09-13 13:43 ` [PATCH 0/4] Optimized generic " H.J. Lu
@ 2017-09-13 13:48   ` Szabolcs Nagy
  0 siblings, 0 replies; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-13 13:48 UTC (permalink / raw)
  To: H.J. Lu; +Cc: nd, GNU C Library

On 13/09/17 14:43, H.J. Lu wrote:
> On Wed, Sep 13, 2017 at 3:43 AM, Szabolcs Nagy <szabolcs.nagy@arm.com> wrote:
>> broke down the expf patch into 4 parts.
>> build-many-glibcs is still running i'll report the
>> results once it's finished.
>>
>> Szabolcs Nagy (4):
>>   Optimized generic expf and exp2f with wrappers
>>   Move exp compat wrappers under math/
>>   New expf and exp2f version without SVID compat wrapper
>>   Do not wrap expf and exp2f
> 
> Do you have a glibc git branch I can try?
> 

i don't have a public branch for this currently, sorry

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-13 10:45 ` [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers Szabolcs Nagy
  2017-09-13 12:41   ` Joseph Myers
@ 2017-09-13 20:13   ` Joseph Myers
  2017-09-14  9:39     ` Szabolcs Nagy
  1 sibling, 1 reply; 19+ messages in thread
From: Joseph Myers @ 2017-09-13 20:13 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: GNU C Library, nd

On Wed, 13 Sep 2017, Szabolcs Nagy wrote:

> - disable math_errf.c and e_exp2f_data.c on i386 and ia64.

sysdeps/m68k/m680x0/fpu has its own e_expf.c and e_exp2f.c, I would expect 
those new files to be disabled there as well.

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 3/4] New expf and exp2f version without SVID compat wrapper
  2017-09-13 10:47 ` [PATCH 3/4] New expf and exp2f version without SVID compat wrapper Szabolcs Nagy
@ 2017-09-13 20:17   ` Joseph Myers
  0 siblings, 0 replies; 19+ messages in thread
From: Joseph Myers @ 2017-09-13 20:17 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: GNU C Library, nd

On Wed, 13 Sep 2017, Szabolcs Nagy wrote:

> This patch changes the expf and exp2f error handling
> semantics to only set errno accoring to POSIX rules.
> New symbol version is introduced at GLIBC_2.27.
> 
> The old wrappers are kept for compat symbols (with
> GLIBC_2_0 version and unchanged error semantics).
> 
> Internal calls to __expf now get the new error semantics,
> this seems to only affect sysdeps/i386/fpu/s_expm1f.S
> where the errno-only behaviour should be correct.
> 
> All linux libm abilists are updated (even ones where i
> could not build a working toolchain).

To confirm: does just adding the versions to Versions like you did 
actually work to get the symbols exported for both versions on ia64 (which 
does not use the common wrappers)?  (I'm fine with having both versions 
point to the same functions there, as long as the versions do really exist 
as indicated in the ABI baseline.)

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 4/4] Do not wrap expf and exp2f
  2017-09-13 10:48 ` [PATCH 4/4] Do not wrap expf and exp2f Szabolcs Nagy
@ 2017-09-13 20:19   ` Joseph Myers
  2017-09-14  0:31     ` Arjan van de Ven
  0 siblings, 1 reply; 19+ messages in thread
From: Joseph Myers @ 2017-09-13 20:19 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: GNU C Library, nd

On Wed, 13 Sep 2017, Szabolcs Nagy wrote:

> 	* sysdeps/x86_64/fpu/w_expf.c: New file.

I don't think having w_expf.c alongside w_expf.S in the same directory 
makes sense.  (An IFUNC w_expf.c in the multiarch subdirectory, used to 
build both fma and non-fma versions and replacing the existing 
e_expf-fma.S, may well make sense if benchmarks show improvements on the 
fma and non-fma .S files.)

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: Re: [PATCH 4/4] Do not wrap expf and exp2f
  2017-09-13 20:19   ` Joseph Myers
@ 2017-09-14  0:31     ` Arjan van de Ven
  0 siblings, 0 replies; 19+ messages in thread
From: Arjan van de Ven @ 2017-09-14  0:31 UTC (permalink / raw)
  To: Joseph Myers, Szabolcs Nagy; +Cc: GNU C Library, nd

On 9/13/2017 1:19 PM, Joseph Myers wrote:
> On Wed, 13 Sep 2017, Szabolcs Nagy wrote:
> 
>> 	* sysdeps/x86_64/fpu/w_expf.c: New file.
> 
> I don't think having w_expf.c alongside w_expf.S in the same directory
> makes sense.  (An IFUNC w_expf.c in the multiarch subdirectory, used to
> build both fma and non-fma versions and replacing the existing
> e_expf-fma.S, may well make sense if benchmarks show improvements on the
> fma and non-fma .S files.)

fwiw I am assuming we'll retire the x86-64 .S files and use the C version
(for both FMA and non-FMA with some IFUNC selector) once this lands.
Optimizing the .S file was fun for FMA, but the C version has a smaller
polynomial and is faster ... data speaks and this wins.


^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-13 20:13   ` Joseph Myers
@ 2017-09-14  9:39     ` Szabolcs Nagy
  2017-09-14 11:48       ` Joseph Myers
  2017-09-14 12:04       ` Szabolcs Nagy
  0 siblings, 2 replies; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-14  9:39 UTC (permalink / raw)
  To: Joseph Myers; +Cc: nd, GNU C Library

On 13/09/17 21:13, Joseph Myers wrote:
> On Wed, 13 Sep 2017, Szabolcs Nagy wrote:
> 
>> - disable math_errf.c and e_exp2f_data.c on i386 and ia64.
> 
> sysdeps/m68k/m680x0/fpu has its own e_expf.c and e_exp2f.c, I would expect 
> those new files to be disabled there as well.
> 

yes i made a number of mistakes:
- m68k/m680x0/fpu does not need the math_errf/e_exp2f_data code
- powerpc64/fpu/multiarch/e_expf-ppc64.c needs to suppress
libm_hidden_proto too now that it overrides __expf
(otherwise there is a hidden proto of __ieee754_expf_ppc64)
- sysdeps/*/w_expf.c does not work to include srcdir/math/w_expf.c
because the generated builddir/math/w_expf.c gets included.
(i'm not yet sure what the right fix for this is, i can just
copy the math/w_expf.c code around instead of including it
or use a file name that does not collide with the generated files)
- the new versioned symbols are missing on ia64 too because
it uses its own w_expf wrappers.

i'm trying to fix these.

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-14  9:39     ` Szabolcs Nagy
@ 2017-09-14 11:48       ` Joseph Myers
  2017-09-14 15:08         ` Szabolcs Nagy
  2017-09-14 12:04       ` Szabolcs Nagy
  1 sibling, 1 reply; 19+ messages in thread
From: Joseph Myers @ 2017-09-14 11:48 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: nd, GNU C Library

On Thu, 14 Sep 2017, Szabolcs Nagy wrote:

> - sysdeps/*/w_expf.c does not work to include srcdir/math/w_expf.c
> because the generated builddir/math/w_expf.c gets included.
> (i'm not yet sure what the right fix for this is, i can just
> copy the math/w_expf.c code around instead of including it
> or use a file name that does not collide with the generated files)

I'd suggest not having a srcdir/math/w_expf.c.  When you need a new 
w_expf.c, put it in sysdeps/ieee754/flt-32 instead.

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-14  9:39     ` Szabolcs Nagy
  2017-09-14 11:48       ` Joseph Myers
@ 2017-09-14 12:04       ` Szabolcs Nagy
  2017-09-14 12:45         ` Joseph Myers
  1 sibling, 1 reply; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-14 12:04 UTC (permalink / raw)
  To: Joseph Myers; +Cc: nd, GNU C Library

On 14/09/17 10:39, Szabolcs Nagy wrote:
> On 13/09/17 21:13, Joseph Myers wrote:
>> On Wed, 13 Sep 2017, Szabolcs Nagy wrote:
>>
>>> - disable math_errf.c and e_exp2f_data.c on i386 and ia64.
>>
>> sysdeps/m68k/m680x0/fpu has its own e_expf.c and e_exp2f.c, I would expect 
>> those new files to be disabled there as well.
>>
> 
> yes i made a number of mistakes:
> - m68k/m680x0/fpu does not need the math_errf/e_exp2f_data code
> - powerpc64/fpu/multiarch/e_expf-ppc64.c needs to suppress
> libm_hidden_proto too now that it overrides __expf
> (otherwise there is a hidden proto of __ieee754_expf_ppc64)
> - sysdeps/*/w_expf.c does not work to include srcdir/math/w_expf.c
> because the generated builddir/math/w_expf.c gets included.
> (i'm not yet sure what the right fix for this is, i can just
> copy the math/w_expf.c code around instead of including it
> or use a file name that does not collide with the generated files)

i guess i can just do
#include <sysdeps/../math/w_expf.c>
to make it clear that i want the file from the srcdir

> - the new versioned symbols are missing on ia64 too because
> it uses its own w_expf wrappers.

i will try to add .symver directives to the ia64
asm files in some ways.

but overall the symbol version fixes need a lot of
changes.. do we really need to carry asm math code
for barely supported targets? do ppl really care
if m68k/ia64/i386 get suboptimal math code?


^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-14 12:04       ` Szabolcs Nagy
@ 2017-09-14 12:45         ` Joseph Myers
  0 siblings, 0 replies; 19+ messages in thread
From: Joseph Myers @ 2017-09-14 12:45 UTC (permalink / raw)
  To: Szabolcs Nagy; +Cc: nd, GNU C Library

On Thu, 14 Sep 2017, Szabolcs Nagy wrote:

> but overall the symbol version fixes need a lot of
> changes.. do we really need to carry asm math code
> for barely supported targets? do ppl really care
> if m68k/ia64/i386 get suboptimal math code?

The first question would be whether it's suboptimal anyway.  That needs 
benchmarking on hardware, which presumably architecture maintainers could 
do for m68k/ia64.  For i386 there are both generic and SSE2 versions of 
expf to benchmark.

It's far from obvious that just because the x87 f2xm1 instruction exists 
it's efficient to use (especially since it needs to produce a result to 
the precision of long double, and expf only needs to produce a result to 
the precision of float).

-- 
Joseph S. Myers
joseph@codesourcery.com

^ permalink raw reply	[flat|nested] 19+ messages in thread

* Re: [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers
  2017-09-14 11:48       ` Joseph Myers
@ 2017-09-14 15:08         ` Szabolcs Nagy
  0 siblings, 0 replies; 19+ messages in thread
From: Szabolcs Nagy @ 2017-09-14 15:08 UTC (permalink / raw)
  To: Joseph Myers; +Cc: nd, GNU C Library

On 14/09/17 12:48, Joseph Myers wrote:
> On Thu, 14 Sep 2017, Szabolcs Nagy wrote:
> 
>> - sysdeps/*/w_expf.c does not work to include srcdir/math/w_expf.c
>> because the generated builddir/math/w_expf.c gets included.
>> (i'm not yet sure what the right fix for this is, i can just
>> copy the math/w_expf.c code around instead of including it
>> or use a file name that does not collide with the generated files)
> 
> I'd suggest not having a srcdir/math/w_expf.c.  When you need a new 
> w_expf.c, put it in sysdeps/ieee754/flt-32 instead.
> 

well eventually i need the w_expf.c at many places
except in sysdeps/ieee754/flt-32 where the generic
code needs no wrapping.

so i have to put it somewhere else such that the
target specific w_expf.c can include it.

^ permalink raw reply	[flat|nested] 19+ messages in thread

end of thread, other threads:[~2017-09-14 15:08 UTC | newest]

Thread overview: 19+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-09-13 10:43 [PATCH 0/4] Optimized generic expf and exp2f Szabolcs Nagy
2017-09-13 10:45 ` [PATCH 1/4 v2] Optimized generic expf and exp2f with wrappers Szabolcs Nagy
2017-09-13 12:41   ` Joseph Myers
2017-09-13 20:13   ` Joseph Myers
2017-09-14  9:39     ` Szabolcs Nagy
2017-09-14 11:48       ` Joseph Myers
2017-09-14 15:08         ` Szabolcs Nagy
2017-09-14 12:04       ` Szabolcs Nagy
2017-09-14 12:45         ` Joseph Myers
2017-09-13 10:46 ` [PATCH 2/4] Move exp compat wrappers under math/ Szabolcs Nagy
2017-09-13 10:49   ` Szabolcs Nagy
2017-09-13 12:43     ` Joseph Myers
2017-09-13 10:47 ` [PATCH 3/4] New expf and exp2f version without SVID compat wrapper Szabolcs Nagy
2017-09-13 20:17   ` Joseph Myers
2017-09-13 10:48 ` [PATCH 4/4] Do not wrap expf and exp2f Szabolcs Nagy
2017-09-13 20:19   ` Joseph Myers
2017-09-14  0:31     ` Arjan van de Ven
2017-09-13 13:43 ` [PATCH 0/4] Optimized generic " H.J. Lu
2017-09-13 13:48   ` Szabolcs Nagy

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).