diff options
| author | H.J. Lu <hjl.tools@gmail.com> | 2017-10-22 07:54:38 -0700 |
|---|---|---|
| committer | H.J. Lu <hjl.tools@gmail.com> | 2017-10-22 07:54:50 -0700 |
| commit | b2f6137ea570933fb4be286574cc58b794eb5b5f (patch) | |
| tree | b95f459a9090aba7854e32c454001a610f400728 /sysdeps/i386/fpu/e_expf.S | |
| parent | e1f59bebd885c442c14df3554da9fe08792ce7ce (diff) | |
| download | glibc-b2f6137ea570933fb4be286574cc58b794eb5b5f.tar glibc-b2f6137ea570933fb4be286574cc58b794eb5b5f.tar.gz glibc-b2f6137ea570933fb4be286574cc58b794eb5b5f.tar.bz2 glibc-b2f6137ea570933fb4be286574cc58b794eb5b5f.zip | |
i386: Replace assembly versions of e_expf with generic e_expf.c
This patch replaces i386 assembly versions of e_expf with generic
e_expf.c. For workload-spec2017.wrf, on Nehalem, it improves
performance by:
Before After Improvement
reciprocal-throughput 55.5724 40.2664 38%
latency 80.0687 60.8517 31%
On Skylake, it improves performance by:
Before After Improvement
reciprocal-throughput 62.4056 39.4188 58%
latency 85.5496 59.6377 43%
On IvyBridge with --disable-multi-arch, it improves performance by:
Before After Improvement
reciprocal-throughput 133.707 40.3778 231%
latency 149.191 63.2515 135%
* sysdeps/i386/fpu/e_exp2f_data.c: Removed.
* sysdeps/i386/fpu/e_expf.S: Likewise.
* sysdeps/i386/fpu/math_errf.c: Likewise.
* sysdeps/i386/fpu/w_expf.c: Likewise.
* sysdeps/i386/i686/fpu/multiarch/e_expf-ia32.S: Likewise.
* sysdeps/i386/i686/fpu/multiarch/e_expf-sse2.S: Likewise.
* sysdeps/i386/i686/fpu/multiarch/w_expf.c: Likewise.
* sysdeps/i386/fpu/libm-test-ulps: Updated for generic e_expf.c.
* sysdeps/i386/i686/fpu/multiarch/libm-test-ulps: Likewise.
* sysdeps/i386/i686/fpu/multiarch/Makefile (libm-sysdep_routines):
Remove e_expf-ia32.
(CFLAGS-e_expf-sse2.c): New.
* sysdeps/i386/i686/fpu/multiarch/e_expf-sse2.c: New file.
* sysdeps/i386/i686/fpu/multiarch/e_expf.c: Rewritten.
Diffstat (limited to 'sysdeps/i386/fpu/e_expf.S')
| -rw-r--r-- | sysdeps/i386/fpu/e_expf.S | 74 |
1 files changed, 0 insertions, 74 deletions
diff --git a/sysdeps/i386/fpu/e_expf.S b/sysdeps/i386/fpu/e_expf.S deleted file mode 100644 index 65cb4ec204..0000000000 --- a/sysdeps/i386/fpu/e_expf.S +++ /dev/null @@ -1,74 +0,0 @@ -/* - * Written by J.T. Conklin <jtc@netbsd.org>. - * Public domain. - * Adapted for float type by Ulrich Drepper <drepper@cygnus.com>. - */ - -#include <machine/asm.h> -#include <i386-math-asm.h> - -DEFINE_FLT_MIN - -#ifdef PIC -# define MO(op) op##@GOTOFF(%ecx) -#else -# define MO(op) op -#endif - - .text -/* e^x = 2^(x * log2(e)) */ -ENTRY(__ieee754_expf) -#ifdef PIC - LOAD_PIC_REG (cx) -#endif - flds 4(%esp) -/* I added the following ugly construct because exp(+-Inf) resulted - in NaN. The ugliness results from the bright minds at Intel. - For the i686 the code can be written better. - -- drepper@cygnus.com. */ - fxam /* Is NaN or +-Inf? */ - fstsw %ax - movb $0x45, %dh - andb %ah, %dh - cmpb $0x05, %dh - je 1f /* Is +-Inf, jump. */ - fldl2e - fmulp /* x * log2(e) */ - fld %st - frndint /* int(x * log2(e)) */ - fsubr %st,%st(1) /* fract(x * log2(e)) */ - fxch - f2xm1 /* 2^(fract(x * log2(e))) - 1 */ - fld1 - faddp /* 2^(fract(x * log2(e))) */ - fscale /* e^x */ - fstp %st(1) - FLT_NARROW_EVAL_UFLOW_NONNEG_NAN - ret - -1: testl $0x200, %eax /* Test sign. */ - jz 2f /* If positive, jump. */ - fstp %st - fldz /* Set result to 0. */ -2: ret -END (__ieee754_expf) - - -ENTRY(__expf_finite) -#ifdef PIC - LOAD_PIC_REG (cx) -#endif - fldl2e - fmuls 4(%esp) /* x * log2(e) */ - fld %st - frndint /* int(x * log2(e)) */ - fsubr %st,%st(1) /* fract(x * log2(e)) */ - fxch - f2xm1 /* 2^(fract(x * log2(e))) - 1 */ - fld1 - faddp /* 2^(fract(x * log2(e))) */ - fscale /* e^x */ - fstp %st(1) - FLT_NARROW_EVAL_UFLOW_NONNEG - ret -END(__expf_finite) |