Optimize accurate 64-bit routines for FMA4 on x86-64

31 files changed, 328 insertions, 0 deletions

diff --git a/sysdeps/x86_64/fpu/multiarch/Makefile b/sysdeps/x86_64/fpu/multiarch/Makefile
index bd07e98..70cb740 100644
--- a/sysdeps/x86_64/fpu/multiarch/Makefile
+++ b/sysdeps/x86_64/fpu/multiarch/Makefile
@@ -1,4 +1,36 @@
 ifeq ($(subdir),math)
 libm-sysdep_routines += s_floor-c s_ceil-c s_floorf-c s_ceilf-c \
 			s_rint-c s_rintf-c s_nearbyint-c s_nearbyintf-c
+
+ifeq ($(have-mfma4),yes)
+libm-sysdep_routines += e_exp-fma4 e_log-fma4 e_pow-fma4 s_atan-fma4 \
+			e_asin-fma4 e_atan2-fma4 s_sin-fma4 s_tan-fma4 \
+			mplog-fma4 mpa-fma4 slowexp-fma4 slowpow-fma4 \
+			sincos32-fma4 doasin-fma4 dosincos-fma4 \
+			brandred-fma4 halfulp-fma4 mpexp-fma4 \
+			mpatan2-fma4 mpatan-fma4 mpsqrt-fma4 mptan-fma4
+
+CFLAGS-brandred-fma4.c = -mfma4
+CFLAGS-doasin-fma4.c = -mfma4
+CFLAGS-dosincos-fma4.c = -mfma4
+CFLAGS-e_asin-fma4.c = -mfma4
+CFLAGS-e_atan2-fma4.c = -mfma4
+CFLAGS-e_exp-fma4.c = -mfma4
+CFLAGS-e_log-fma4.c = -mfma4
+CFLAGS-e_pow-fma4.c = -mfma4
+CFLAGS-halfulp-fma4.c = -mfma4
+CFLAGS-mpa-fma4.c = -mfma4
+CFLAGS-mpatan-fma4.c = -mfma4
+CFLAGS-mpatan2-fma4.c = -mfma4
+CFLAGS-mpexp-fma4.c = -mfma4
+CFLAGS-mplog-fma4.c = -mfma4
+CFLAGS-mpsqrt-fma4.c = -mfma4
+CFLAGS-mptan-fma4.c = -mfma4
+CFLAGS-s_atan-fma4.c = -mfma4
+CFLAGS-sincos32-fma4.c = -mfma4
+CFLAGS-slowexp-fma4.c = -mfma4
+CFLAGS-slowpow-fma4.c = -mfma4
+CLFAGS-s_sin-fma4.c = -mfma4
+CLFAGS-s_tan-fma4.c = -mfma4
+endif
 endif
diff --git a/sysdeps/x86_64/fpu/multiarch/brandred-fma4.c b/sysdeps/x86_64/fpu/multiarch/brandred-fma4.c
new file mode 100644
index 0000000..93fb5a1
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/brandred-fma4.c
@@ -0,0 +1,3 @@
+#define __branred __branred_fma4
+
+#include <sysdeps/ieee754/dbl-64/branred.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/doasin-fma4.c b/sysdeps/x86_64/fpu/multiarch/doasin-fma4.c
new file mode 100644
index 0000000..d7ba67e
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/doasin-fma4.c
@@ -0,0 +1,3 @@
+#define __doasin __doasin_fma4
+
+#include <sysdeps/ieee754/dbl-64/doasin.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/dosincos-fma4.c b/sysdeps/x86_64/fpu/multiarch/dosincos-fma4.c
new file mode 100644
index 0000000..02b420b
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/dosincos-fma4.c
@@ -0,0 +1,5 @@
+#define __docos __docos_fma4
+#define __dubcos __dubcos_fma4
+#define __dubsin __dubsin_fma4
+
+#include <sysdeps/ieee754/dbl-64/dosincos.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_asin-fma4.c b/sysdeps/x86_64/fpu/multiarch/e_asin-fma4.c
new file mode 100644
index 0000000..938bc84
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_asin-fma4.c
@@ -0,0 +1,10 @@
+#define __ieee754_acos __ieee754_acos_fma4
+#define __ieee754_asin __ieee754_asin_fma4
+#define __cos32 __cos32_fma4
+#define __doasin __doasin_fma4
+#define __docos __docos_fma4
+#define __dubcos __dubcos_fma4
+#define __dubsin __dubsin_fma4
+#define __sin32 __sin32_fma4
+
+#include <sysdeps/ieee754/dbl-64/e_asin.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_asin.c b/sysdeps/x86_64/fpu/multiarch/e_asin.c
new file mode 100644
index 0000000..8882cea
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_asin.c
@@ -0,0 +1,23 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math_private.h>
+
+extern double __ieee754_acos_sse2 (double);
+extern double __ieee754_acos_fma4 (double);
+extern double __ieee754_asin_sse2 (double);
+extern double __ieee754_asin_fma4 (double);
+
+libm_ifunc (__ieee754_acos,
+	    HAS_FMA4 ? __ieee754_acos_fma4 : __ieee754_acos_sse2);
+strong_alias (__ieee754_acos, __acos_finite)
+
+libm_ifunc (__ieee754_asin,
+	    HAS_FMA4 ? __ieee754_asin_fma4 : __ieee754_asin_sse2);
+strong_alias (__ieee754_asin, __asin_finite)
+
+# define __ieee754_acos __ieee754_acos_sse2
+# define __ieee754_asin __ieee754_asin_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/e_asin.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_atan2-fma4.c b/sysdeps/x86_64/fpu/multiarch/e_atan2-fma4.c
new file mode 100644
index 0000000..84a6f86
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_atan2-fma4.c
@@ -0,0 +1,9 @@
+#define __ieee754_atan2 __ieee754_atan2_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __dvd __dvd_fma4
+#define __mpatan2 __mpatan2_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/e_atan2.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_atan2.c b/sysdeps/x86_64/fpu/multiarch/e_atan2.c
new file mode 100644
index 0000000..12fc929
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_atan2.c
@@ -0,0 +1,16 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math_private.h>
+
+extern double __ieee754_atan2_sse2 (double, double);
+extern double __ieee754_atan2_fma4 (double, double);
+
+libm_ifunc (__ieee754_atan2,
+	    HAS_FMA4 ? __ieee754_atan2_fma4 : __ieee754_atan2_sse2);
+strong_alias (__ieee754_atan2, __atan2_finite)
+
+# define __ieee754_atan2 __ieee754_atan2_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/e_atan2.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_exp-fma4.c b/sysdeps/x86_64/fpu/multiarch/e_exp-fma4.c
new file mode 100644
index 0000000..942dfff
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_exp-fma4.c
@@ -0,0 +1,5 @@
+#define __ieee754_exp __ieee754_exp_fma4
+#define __exp1 __exp1_fma4
+#define __slowexp __slowexp_fma4
+
+#include <sysdeps/ieee754/dbl-64/e_exp.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_exp.c b/sysdeps/x86_64/fpu/multiarch/e_exp.c
new file mode 100644
index 0000000..fc1096b
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_exp.c
@@ -0,0 +1,15 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math_private.h>
+
+extern double __ieee754_exp_sse2 (double);
+extern double __ieee754_exp_fma4 (double);
+
+libm_ifunc (__ieee754_exp, HAS_FMA4 ? __ieee754_exp_fma4 : __ieee754_exp_sse2);
+strong_alias (__ieee754_exp, __exp_finite)
+
+# define __ieee754_exp __ieee754_exp_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/e_exp.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_log-fma4.c b/sysdeps/x86_64/fpu/multiarch/e_log-fma4.c
new file mode 100644
index 0000000..0be66d1
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_log-fma4.c
@@ -0,0 +1,7 @@
+#define __ieee754_log __ieee754_log_fma4
+#define __mplog __mplog_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/e_log.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_log.c b/sysdeps/x86_64/fpu/multiarch/e_log.c
new file mode 100644
index 0000000..c542646
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_log.c
@@ -0,0 +1,15 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math_private.h>
+
+extern double __ieee754_log_sse2 (double);
+extern double __ieee754_log_fma4 (double);
+
+libm_ifunc (__ieee754_log, HAS_FMA4 ? __ieee754_log_fma4 : __ieee754_log_sse2);
+strong_alias (__ieee754_log, __log_finite)
+
+# define __ieee754_log __ieee754_log_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/e_log.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_pow-fma4.c b/sysdeps/x86_64/fpu/multiarch/e_pow-fma4.c
new file mode 100644
index 0000000..20313be
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_pow-fma4.c
@@ -0,0 +1,5 @@
+#define __ieee754_pow __ieee754_pow_fma4
+#define __exp1 __exp1_fma4
+#define __slowpow __slowpow_fma4
+
+#include <sysdeps/ieee754/dbl-64/e_pow.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_pow.c b/sysdeps/x86_64/fpu/multiarch/e_pow.c
new file mode 100644
index 0000000..a740b6c
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_pow.c
@@ -0,0 +1,15 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math_private.h>
+
+extern double __ieee754_pow_sse2 (double, double);
+extern double __ieee754_pow_fma4 (double, double);
+
+libm_ifunc (__ieee754_pow, HAS_FMA4 ? __ieee754_pow_fma4 : __ieee754_pow_sse2);
+strong_alias (__ieee754_pow, __pow_finite)
+
+# define __ieee754_pow __ieee754_pow_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/e_pow.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/halfulp-fma4.c b/sysdeps/x86_64/fpu/multiarch/halfulp-fma4.c
new file mode 100644
index 0000000..3fc223e
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/halfulp-fma4.c
@@ -0,0 +1,3 @@
+#define __halfulp __halfulp_fma4
+
+#include <sysdeps/ieee754/dbl-64/halfulp.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mpa-fma4.c b/sysdeps/x86_64/fpu/multiarch/mpa-fma4.c
new file mode 100644
index 0000000..7b9e2ef
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mpa-fma4.c
@@ -0,0 +1,10 @@
+#define __add __add_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __dvd __dvd_fma4
+
+#define NO___CPY 1
+#define NO___MP_DBL 1
+
+#include <sysdeps/ieee754/dbl-64/mpa.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mpatan-fma4.c b/sysdeps/x86_64/fpu/multiarch/mpatan-fma4.c
new file mode 100644
index 0000000..942974b
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mpatan-fma4.c
@@ -0,0 +1,8 @@
+#define __mpatan __mpatan_fma4
+#define __add __add_fma4
+#define __dvd __dvd_fma4
+#define __mpsqrt __mpsqrt_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/mpatan.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mpatan2-fma4.c b/sysdeps/x86_64/fpu/multiarch/mpatan2-fma4.c
new file mode 100644
index 0000000..e7c469e
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mpatan2-fma4.c
@@ -0,0 +1,8 @@
+#define __mpatan2 __mpatan2_fma4
+#define __add __add_fma4
+#define __dvd __dvd_fma4
+#define __mpatan __mpatan_fma4
+#define __mpsqrt __mpsqrt_fma4
+#define __mul __mul_fma4
+
+#include <sysdeps/ieee754/dbl-64/mpatan2.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mpexp-fma4.c b/sysdeps/x86_64/fpu/multiarch/mpexp-fma4.c
new file mode 100644
index 0000000..021970c
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mpexp-fma4.c
@@ -0,0 +1,7 @@
+#define __mpexp __mpexp_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __dvd __dvd_fma4
+#define __mul __mul_fma4
+
+#include <sysdeps/ieee754/dbl-64/mpexp.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mplog-fma4.c b/sysdeps/x86_64/fpu/multiarch/mplog-fma4.c
new file mode 100644
index 0000000..9581eaf
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mplog-fma4.c
@@ -0,0 +1,7 @@
+#define __mplog __mplog_fma4
+#define __add __add_fma4
+#define __mpexp __mpexp_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/mplog.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mpsqrt-fma4.c b/sysdeps/x86_64/fpu/multiarch/mpsqrt-fma4.c
new file mode 100644
index 0000000..43b6493
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mpsqrt-fma4.c
@@ -0,0 +1,6 @@
+#define __mpsqrt __mpsqrt_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/mpsqrt.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/mptan-fma4.c b/sysdeps/x86_64/fpu/multiarch/mptan-fma4.c
new file mode 100644
index 0000000..767924e
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/mptan-fma4.c
@@ -0,0 +1,6 @@
+#define __mptan __mptan_fma4
+#define __c32 __c32_fma4
+#define __dvd __dvd_fma4
+#define __mpranred __mpranred_fma4
+
+#include <sysdeps/ieee754/dbl-64/mptan.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/s_atan-fma4.c b/sysdeps/x86_64/fpu/multiarch/s_atan-fma4.c
new file mode 100644
index 0000000..a8f0977
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/s_atan-fma4.c
@@ -0,0 +1,8 @@
+#define atan __atan_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __mpatan __mpatan_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/s_atan.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/s_atan.c b/sysdeps/x86_64/fpu/multiarch/s_atan.c
new file mode 100644
index 0000000..ffc4a56
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/s_atan.c
@@ -0,0 +1,14 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math.h>
+
+extern double __atan_sse2 (double);
+extern double __atan_fma4 (double);
+
+libm_ifunc (atan, HAS_FMA4 ? __atan_fma4 : __atan_sse2);
+
+# define atan __atan_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/s_atan.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/s_sin-fma4.c b/sysdeps/x86_64/fpu/multiarch/s_sin-fma4.c
new file mode 100644
index 0000000..97cef8b
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/s_sin-fma4.c
@@ -0,0 +1,11 @@
+#define __cos __cos_fma4
+#define __sin __sin_fma4
+#define __branred __branred_fma4
+#define __docos __docos_fma4
+#define __dubsin __dubsin_fma4
+#define __mpcos __mpcos_fma4
+#define __mpcos1 __mpcos1_fma4
+#define __mpsin __mpsin_fma4
+#define __mpsin1 __mpsin1_fma4
+
+#include <sysdeps/ieee754/dbl-64/s_sin.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/s_sin.c b/sysdeps/x86_64/fpu/multiarch/s_sin.c
new file mode 100644
index 0000000..a7c35dc
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/s_sin.c
@@ -0,0 +1,22 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math.h>
+# undef NAN
+
+extern double __cos_sse2 (double);
+extern double __cos_fma4 (double);
+extern double __sin_sse2 (double);
+extern double __sin_fma4 (double);
+
+libm_ifunc (__cos, HAS_FMA4 ? __cos_fma4 : __cos_sse2);
+weak_alias (__cos, cos)
+
+libm_ifunc (__sin, HAS_FMA4 ? __sin_fma4 : __sin_sse2);
+weak_alias (__sin, sin)
+
+# define __cos __cos_sse2
+# define __sin __sin_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/s_sin.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/s_tan-fma4.c b/sysdeps/x86_64/fpu/multiarch/s_tan-fma4.c
new file mode 100644
index 0000000..c3cefc2
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/s_tan-fma4.c
@@ -0,0 +1,9 @@
+#define tan __tan_fma4
+#define __branred __branred_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __mpranred __mpranred_fma4
+#define __mptan __mptan_fma4
+#define __sub __sub_fma4
+
+
+#include <sysdeps/ieee754/dbl-64/s_tan.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/s_tan.c b/sysdeps/x86_64/fpu/multiarch/s_tan.c
new file mode 100644
index 0000000..cca02b5
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/s_tan.c
@@ -0,0 +1,14 @@
+#ifdef HAVE_FMA4_SUPPORT
+# include <init-arch.h>
+# include <math.h>
+
+extern double __tan_sse2 (double);
+extern double __tan_fma4 (double);
+
+libm_ifunc (tan, HAS_FMA4 ? __tan_fma4 : __tan_sse2);
+
+# define tan __tan_sse2
+#endif
+
+
+#include <sysdeps/ieee754/dbl-64/s_tan.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/sincos32-fma4.c b/sysdeps/x86_64/fpu/multiarch/sincos32-fma4.c
new file mode 100644
index 0000000..f0d2d27
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/sincos32-fma4.c
@@ -0,0 +1,14 @@
+#define __cos32 __cos32_fma4
+#define __sin32 __sin32_fma4
+#define __c32 __c32_fma4
+#define __mpsin __mpsin_fma4
+#define __mpsin1 __mpsin1_fma4
+#define __mpcos __mpcos_fma4
+#define __mpcos1 __mpcos1_fma4
+#define __mpranred __mpranred_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/sincos32.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/slowexp-fma4.c b/sysdeps/x86_64/fpu/multiarch/slowexp-fma4.c
new file mode 100644
index 0000000..83cb359
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/slowexp-fma4.c
@@ -0,0 +1,8 @@
+#define __slowexp __slowexp_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __mpexp __mpexp_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+
+#include <sysdeps/ieee754/dbl-64/slowexp.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/slowpow-fma4.c b/sysdeps/x86_64/fpu/multiarch/slowpow-fma4.c
new file mode 100644
index 0000000..744f3f6
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/slowpow-fma4.c
@@ -0,0 +1,10 @@
+#define __slowpow __slowpow_fma4
+#define __add __add_fma4
+#define __dbl_mp __dbl_mp_fma4
+#define __mpexp __mpexp_fma4
+#define __mplog __mplog_fma4
+#define __mul __mul_fma4
+#define __sub __sub_fma4
+#define __halfulp __halfulp_fma4
+
+#include <sysdeps/ieee754/dbl-64/slowpow.c>