Optimize x86-32 log

7 files changed, 306 insertions, 3 deletions

diff --git a/ChangeLog b/ChangeLog
index 499050b..e2d6f38 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,12 @@
 2011-10-14  Ulrich Drepper  <drepper@gmail.com>
 
+	* sysdeps/i386/fpu/e_log.S: Add real definition of __log_finite.
+	* sysdeps/i386/fpu/e_logf.S: Add real definition of __logf_finite.
+	* sysdeps/i386/fpu/e_logl.S: Add real definition of __logl_finite.
+	* sysdeps/i386/i686/fpu/e_log.S: New file.
+	* sysdeps/i386/i686/fpu/e_logf.S: New file.
+	* sysdeps/i386/i686/fpu/e_logl.S: New file.
+
 	* ctype/ctype.h: Add support for inlined isXXX functions when
 	compiling C++ code.
 
diff --git a/sysdeps/i386/fpu/e_log.S b/sysdeps/i386/fpu/e_log.S
index 8110a84..a2e4d89 100644
--- a/sysdeps/i386/fpu/e_log.S
+++ b/sysdeps/i386/fpu/e_log.S
@@ -62,4 +62,22 @@ ENTRY(__ieee754_log)
 	fstp	%st(1)
 	ret
 END (__ieee754_log)
-strong_alias (__ieee754_log, __log_finite)
+
+ENTRY(__log_finite)
+	fldln2			// log(2)
+	fldl	4(%esp)		// x : log(2)
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fcompl	MO(limit)	// x-1 : x : log(2)
+	fnstsw			// x-1 : x : log(2)
+	andb	$0x45, %ah
+	jz	2b
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+END(__log_finite)
diff --git a/sysdeps/i386/fpu/e_logf.S b/sysdeps/i386/fpu/e_logf.S
index b683e13..1992cc2 100644
--- a/sysdeps/i386/fpu/e_logf.S
+++ b/sysdeps/i386/fpu/e_logf.S
@@ -63,4 +63,22 @@ ENTRY(__ieee754_logf)
 	fstp	%st(1)
 	ret
 END (__ieee754_logf)
-strong_alias (__ieee754_logf, __logf_finite)
+
+ENTRY(__logf_finite)
+	fldln2			// log(2)
+	flds	4(%esp)		// x : log(2)
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fcompl	MO(limit)	// x-1 : x : log(2)
+	fnstsw			// x-1 : x : log(2)
+	andb	$0x45, %ah
+	jz	2b
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+END(__logf_finite)
diff --git a/sysdeps/i386/fpu/e_logl.S b/sysdeps/i386/fpu/e_logl.S
index ee1fb16..bfb72a3 100644
--- a/sysdeps/i386/fpu/e_logl.S
+++ b/sysdeps/i386/fpu/e_logl.S
@@ -63,4 +63,22 @@ ENTRY(__ieee754_logl)
 	fstp	%st(1)
 	ret
 END (__ieee754_logl)
-strong_alias (__ieee754_logl, __logl_finite)
+
+ENTRY(__logl_finite)
+	fldln2			// log(2)
+	fldt	4(%esp)		// x : log(2)
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fcompl	MO(limit)	// x-1 : x : log(2)
+	fnstsw			// x-1 : x : log(2)
+	andb	$0x45, %ah
+	jz	2b
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+END(__logl_finite)
diff --git a/sysdeps/i386/i686/fpu/e_log.S b/sysdeps/i386/i686/fpu/e_log.S
new file mode 100644
index 0000000..c6524b1
--- /dev/null
+++ b/sysdeps/i386/i686/fpu/e_log.S
@@ -0,0 +1,80 @@
+/*
+ * Written by J.T. Conklin <jtc@netbsd.org>.
+ * Public domain.
+ *
+ * Changed to use fyl2xp1 for values near 1, <drepper@cygnus.com>.
+ * Adapted for i686 instructions.
+ */
+
+#include <machine/asm.h>
+
+#ifdef __ELF__
+	.section .rodata.cst8,"aM",@progbits,8
+#else
+	.text
+#endif
+	.p2align 3
+	ASM_TYPE_DIRECTIVE(one,@object)
+one:	.double 1.0
+	ASM_SIZE_DIRECTIVE(one)
+	/* It is not important that this constant is precise.  It is only
+	   a value which is known to be on the safe side for using the
+	   fyl2xp1 instruction.  */
+	ASM_TYPE_DIRECTIVE(limit,@object)
+limit:	.double 0.29
+	ASM_SIZE_DIRECTIVE(limit)
+
+
+#ifdef PIC
+# define MO(op) op##@GOTOFF(%edx)
+#else
+# define MO(op) op
+#endif
+
+	.text
+ENTRY(__ieee754_log)
+	fldln2			// log(2)
+	fldl	4(%esp)		// x : log(2)
+	fucomi	%st
+	jp	3f
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fld	MO(limit)	// 0.29 : |x-1| : x-1 : x : log(2)
+	fcomip	%st(1)		// |x-1| : x-1 : x : log(2)
+	fstp	%st(0)		// x-1 : x : log(2)
+	jc	2f
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+
+2:	fstp	%st(0)		// x : log(2)
+	fyl2x			// log(x)
+	ret
+
+3:	fstp	%st(1)
+	ret
+END (__ieee754_log)
+
+ENTRY(__log_finite)
+	fldln2			// log(2)
+	fldl	4(%esp)		// x : log(2)
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fld	MO(limit)	// 0.29 : |x-1| : x-1 : x : log(2)
+	fcomip	%st(1)		// |x-1| : x-1 : x : log(2)
+	fstp	%st(0)		// x-1 : x : log(2)
+	jc	2b
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+END(__log_finite)
diff --git a/sysdeps/i386/i686/fpu/e_logf.S b/sysdeps/i386/i686/fpu/e_logf.S
new file mode 100644
index 0000000..64f8807
--- /dev/null
+++ b/sysdeps/i386/i686/fpu/e_logf.S
@@ -0,0 +1,81 @@
+/*
+ * Written by J.T. Conklin <jtc@netbsd.org>.
+ * Public domain.
+ * Adapted for float by Ulrich Drepper <drepper@cygnus.com>.
+ *
+ * Changed to use fyl2xp1 for values near 1, <drepper@cygnus.com>.
+ * Adapted for i686 instructions.
+ */
+
+#include <machine/asm.h>
+
+#ifdef __ELF__
+	.section .rodata.cst8,"aM",@progbits,8
+#else
+	.text
+#endif
+	.p2align 3
+	ASM_TYPE_DIRECTIVE(one,@object)
+one:	.double 1.0
+	ASM_SIZE_DIRECTIVE(one)
+	/* It is not important that this constant is precise.  It is only
+	   a value which is known to be on the safe side for using the
+	   fyl2xp1 instruction.  */
+	ASM_TYPE_DIRECTIVE(limit,@object)
+limit:	.double 0.29
+	ASM_SIZE_DIRECTIVE(limit)
+
+
+#ifdef PIC
+# define MO(op) op##@GOTOFF(%edx)
+#else
+# define MO(op) op
+#endif
+
+	.text
+ENTRY(__ieee754_logf)
+	fldln2			// log(2)
+	flds	4(%esp)		// x : log(2)
+	fucomi	%st
+	jp	3f
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fld	MO(limit)	// 0.29 : |x-1| : x-1 : x : log(2)
+	fcomip	%st(1)		// |x-1| : x-1 : x : log(2)
+	fstp	%st(0)		// x-1 : x : log(2)
+	jc	2f
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+
+2:	fstp	%st(0)		// x : log(2)
+	fyl2x			// log(x)
+	ret
+
+3:	fstp	%st(1)
+	ret
+END (__ieee754_logf)
+
+ENTRY(__logf_finite)
+	fldln2			// log(2)
+	flds	4(%esp)		// x : log(2)
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fld	MO(limit)	// 0.29 : |x-1| : x-1 : x : log(2)
+	fcomip	%st(1)		// |x-1| : x-1 : x : log(2)
+	fstp	%st(0)		// x-1 : x : log(2)
+	jc	2b
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+END(__logf_finite)
diff --git a/sysdeps/i386/i686/fpu/e_logl.S b/sysdeps/i386/i686/fpu/e_logl.S
new file mode 100644
index 0000000..4e79a5a
--- /dev/null
+++ b/sysdeps/i386/i686/fpu/e_logl.S
@@ -0,0 +1,81 @@
+/*
+ * Written by J.T. Conklin <jtc@netbsd.org>.
+ * Public domain.
+ *
+ * Adapted for `long double' by Ulrich Drepper <drepper@cygnus.com>.
+ * Changed to use fyl2xp1 for values near 1, <drepper@cygnus.com>.
+ * Adapted for i686 instructions.
+ */
+
+#include <machine/asm.h>
+
+#ifdef __ELF__
+	.section .rodata.cst8,"aM",@progbits,8
+#else
+	.text
+#endif
+	.p2align 3
+	ASM_TYPE_DIRECTIVE(one,@object)
+one:	.double 1.0
+	ASM_SIZE_DIRECTIVE(one)
+	/* It is not important that this constant is precise.  It is only
+	   a value which is known to be on the safe side for using the
+	   fyl2xp1 instruction.  */
+	ASM_TYPE_DIRECTIVE(limit,@object)
+limit:	.double 0.29
+	ASM_SIZE_DIRECTIVE(limit)
+
+
+#ifdef PIC
+# define MO(op) op##@GOTOFF(%edx)
+#else
+# define MO(op) op
+#endif
+
+	.text
+ENTRY(__ieee754_logl)
+	fldln2			// log(2)
+	fldt	4(%esp)		// x : log(2)
+	fucomi	%st
+	jp	3f
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fld	MO(limit)	// 0.29 : |x-1| : x-1 : x : log(2)
+	fcomip	%st(1)		// |x-1| : x-1 : x : log(2)
+	fstp	%st(0)		// x-1 : x : log(2)
+	jc	2f
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+
+2:	fstp	%st(0)		// x : log(2)
+	fyl2x			// log(x)
+	ret
+
+3:	fstp	%st(1)
+	ret
+END (__ieee754_logl)
+
+ENTRY(__logl_finite)
+	fldln2			// log(2)
+	fldt	4(%esp)		// x : log(2)
+#ifdef PIC
+	LOAD_PIC_REG (dx)
+#endif
+	fld	%st		// x : x : log(2)
+	fsubl	MO(one)		// x-1 : x : log(2)
+	fld	%st		// x-1 : x-1 : x : log(2)
+	fabs			// |x-1| : x-1 : x : log(2)
+	fld	MO(limit)	// 0.29 : |x-1| : x-1 : x : log(2)
+	fcomip	%st(1)		// |x-1| : x-1 : x : log(2)
+	fstp	%st(0)		// x-1 : x : log(2)
+	jc	2b
+	fstp	%st(1)		// x-1 : log(2)
+	fyl2xp1			// log(x)
+	ret
+END(__logl_finite)