ix87 assembler implementation of cbrt

3 files changed, 608 insertions, 0 deletions

diff --git a/sysdeps/libm-i387/s_cbrt.S b/sysdeps/libm-i387/s_cbrt.S
new file mode 100644
index 0000000000..3f6a0174f2
--- /dev/null
+++ b/sysdeps/libm-i387/s_cbrt.S
@@ -0,0 +1,201 @@
+/* Compute cubic root of double value.
+   Copyright (C) 1997 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Dirk Alboth <dirka@uni-paderborn.de> and
+   Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 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
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <machine/asm.h>
+
+#ifdef __ELF__
+        .section .rodata
+#else
+        .text
+#endif
+
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f7,@object)
+f7:	.double -0.145263899385486377
+	ASM_SIZE_DIRECTIVE(f7)
+        ASM_TYPE_DIRECTIVE(f6,@object)
+f6:	.double 0.784932344976639262
+	ASM_SIZE_DIRECTIVE(f6)
+        ASM_TYPE_DIRECTIVE(f5,@object)
+f5:	.double -1.83469277483613086
+	ASM_SIZE_DIRECTIVE(f5)
+        ASM_TYPE_DIRECTIVE(f4,@object)
+f4:	.double 2.44693122563534430
+	ASM_SIZE_DIRECTIVE(f4)
+        ASM_TYPE_DIRECTIVE(f3,@object)
+f3:	.double -2.11499494167371287
+	ASM_SIZE_DIRECTIVE(f3)
+        ASM_TYPE_DIRECTIVE(f2,@object)
+f2:	.double 1.50819193781584896
+	ASM_SIZE_DIRECTIVE(f2)
+        ASM_TYPE_DIRECTIVE(f1,@object)
+f1:	.double 0.354895765043919860
+	ASM_SIZE_DIRECTIVE(f1)
+
+#define CBRT2		1.2599210498948731648
+#define ONE_CBRT2	0.793700525984099737355196796584
+#define SQR_CBRT2	1.5874010519681994748
+#define ONE_SQR_CBRT2	0.629960524947436582364439673883
+
+	ASM_TYPE_DIRECTIVE(factor,@object)
+factor:	.double ONE_SQR_CBRT2
+	.double ONE_CBRT2
+	.double 1.0
+	.double CBRT2
+	.double SQR_CBRT2
+	ASM_SIZE_DIRECTIVE(factor)
+
+        ASM_TYPE_DIRECTIVE(two54,@object)
+two54:  .byte 0, 0, 0, 0, 0, 0, 0x50, 0x43
+        ASM_SIZE_DIRECTIVE(two54)
+
+#ifdef PIC
+#define MO(op) op##@GOTOFF(%ebx)
+#define MOX(op,x) op##@GOTOFF(%ebx,x,1)
+#else
+#define MO(op) op
+#define MOX(op,x) op(x)
+#endif
+
+	.text
+ENTRY(__cbrt)
+	movl	4(%esp), %ecx
+	movl	8(%esp), %eax
+	movl	%eax, %edx
+	andl	$0x7fffffff, %eax
+	orl	%eax, %ecx
+	jz	1f
+	xorl	%ecx, %ecx
+	cmpl	$0x7ff00000, %eax
+	jae	1f
+
+#ifdef PIC
+	pushl	%ebx
+	call	3f
+3:	popl	%ebx
+	addl    $_GLOBAL_OFFSET_TABLE_+[.-3b], %ebx
+#endif
+
+	cmpl	$0x00100000, %eax
+	jae	2f
+
+#ifdef PIC
+	fldl	8(%esp)
+#else
+	fldl	4(%esp)
+#endif
+	fmull	MO(two54)
+	movl	$-54, %ecx
+#ifdef PIC
+	fstpl	8(%esp)
+	movl	12(%esp), %eax
+#else
+	fstpl	4(%esp)
+	movl	8(%esp), %eax
+#endif
+	movl	%eax, %edx
+	andl	$0x7fffffff, %eax
+
+2:	shrl	$20, %eax
+	andl	$0x800fffff, %edx
+	subl	$1022, %eax
+	orl	$0x3fe00000, %edx
+	addl	%eax, %ecx
+#ifdef PIC
+	movl	%edx, 12(%esp)
+
+	fldl	8(%esp)			/* xm */
+#else
+	movl	%edx, 8(%esp)
+
+	fldl	4(%esp)			/* xm */
+#endif
+	fabs
+
+	/* The following code has two tracks:
+	    a) compute the normalized cbrt value
+	    b) compute xe/3 and xe%3
+	   The right track computes the value for b) and this is done
+	   in an optimized way by avoiding division.
+
+	   But why two tracks at all?  Very easy: efficiency.  Some FP
+	   instruction can overlap with a certain amount of integer (and
+	   FP) instructions.  So we get (except for the imull) all
+	   instructions for free.  */
+
+	fld	%st(0)			/* xm : xm */
+
+	fmull	MO(f7)			/* f7*xm : xm */
+			movl	$1431655766, %eax
+	faddl	MO(f6)			/* f6+f7*xm : xm */
+			imull	%ecx
+	fmul	%st(1)			/* (f6+f7*xm)*xm : xm */
+			movl	%ecx, %eax
+	faddl	MO(f5)			/* f5+(f6+f7*xm)*xm : xm */
+			sarl	$31, %eax
+	fmul	%st(1)			/* (f5+(f6+f7*xm)*xm)*xm : xm */
+			subl	%eax, %edx
+	faddl	MO(f4)			/* f4+(f5+(f6+f7*xm)*xm)*xm : xm */
+	fmul	%st(1)			/* (f4+(f5+(f6+f7*xm)*xm)*xm)*xm : xm */
+	faddl	MO(f3)			/* f3+(f4+(f5+(f6+f7*xm)*xm)*xm)*xm : xm */
+	fmul	%st(1)			/* (f3+(f4+(f5+(f6+f7*xm)*xm)*xm)*xm)*xm : xm */
+	faddl	MO(f2)			/* f2+(f3+(f4+(f5+(f6+f7*xm)*xm)*xm)*xm)*xm : xm */
+	fmul	%st(1)			/* (f2+(f3+(f4+(f5+(f6+f7*xm)*xm)*xm)*xm)*xm)*xm : xm */
+	faddl	MO(f1)			/* u:=f1+(f2+(f3+(f4+(f5+(f6+f7*xm)*xm)*xm)*xm)*xm)*xm : xm */
+
+	fld	%st			/* u : u : xm */
+	fmul	%st(1)			/* u*u : u : xm */
+	fld	%st(2)			/* xm : u*u : u : xm */
+	fadd	%st			/* 2*xm : u*u : u : xm */
+	fxch	%st(1)			/* u*u : 2*xm : u : xm */
+	fmul	%st(2)			/* t2:=u*u*u : 2*xm : u : xm */
+			movl	%edx, %eax
+	fadd	%st, %st(1)		/* t2 : t2+2*xm : u : xm */
+			leal	(%edx,%edx,2),%edx
+	fadd	%st(0)			/* 2*t2 : t2+2*xm : u : xm */
+			subl	%edx, %ecx
+	faddp	%st, %st(3)		/* t2+2*xm : u : 2*t2+xm */
+			shll	$3, %ecx
+	fmulp				/* u*(t2+2*xm) : 2*t2+xm */
+	fdivp	%st, %st(1)		/* u*(t2+2*xm)/(2*t2+xm) */
+	fmull	MOX(16+factor,%ecx)	/* u*(t2+2*xm)/(2*t2+xm)*FACT */
+	pushl	%eax
+	fildl	(%esp)			/* xe/3 : u*(t2+2*xm)/(2*t2+xm)*FACT */
+	fxch				/* u*(t2+2*xm)/(2*t2+xm)*FACT : xe/3 */
+	fscale				/* u*(t2+2*xm)/(2*t2+xm)*FACT*2^xe/3 */
+	popl	%edx
+#ifdef PIC
+	movl	12(%esp), %eax
+	popl	%ebx
+#else
+	movl	8(%esp), %eax
+#endif
+	testl	%eax, %eax
+	fstp	%st(1)
+	jns	4f
+	fchs
+4:	ret
+
+	/* Return the argument.  */
+1:	fldl	4(%esp)
+	ret
+END(__cbrt)
+weak_alias (__cbrt, cbrt)
diff --git a/sysdeps/libm-i387/s_cbrtf.S b/sysdeps/libm-i387/s_cbrtf.S
new file mode 100644
index 0000000000..a14e04ed2f
--- /dev/null
+++ b/sysdeps/libm-i387/s_cbrtf.S
@@ -0,0 +1,178 @@
+/* Compute cubic root of float value.
+   Copyright (C) 1997 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Dirk Alboth <dirka@uni-paderborn.de> and
+   Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 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
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <machine/asm.h>
+
+#ifdef __ELF__
+        .section .rodata
+#else
+        .text
+#endif
+
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f3,@object)
+f3:	.double 0.191502161678719066
+        ASM_SIZE_DIRECTIVE(f3)
+        ASM_TYPE_DIRECTIVE(f2,@object)
+f2:	.double 0.697570460207922770
+        ASM_SIZE_DIRECTIVE(f2)
+        ASM_TYPE_DIRECTIVE(f1,@object)
+f1:	.double 0.492659620528969547
+        ASM_SIZE_DIRECTIVE(f1)
+
+#define CBRT2		1.2599210498948731648
+#define ONE_CBRT2	0.793700525984099737355196796584
+#define SQR_CBRT2	1.5874010519681994748
+#define ONE_SQR_CBRT2	0.629960524947436582364439673883
+
+	ASM_TYPE_DIRECTIVE(factor,@object)
+        .align ALIGNARG(4)
+factor:	.double ONE_SQR_CBRT2
+	.double ONE_CBRT2
+	.double 1.0
+	.double CBRT2
+	.double SQR_CBRT2
+	ASM_SIZE_DIRECTIVE(factor)
+
+        ASM_TYPE_DIRECTIVE(two25,@object)
+two25:	.byte 0, 0, 0, 0x4c
+        ASM_SIZE_DIRECTIVE(two25)
+
+#ifdef PIC
+#define MO(op) op##@GOTOFF(%ebx)
+#define MOX(op,x) op##@GOTOFF(%ebx,x,1)
+#else
+#define MO(op) op
+#define MOX(op,x) op(x)
+#endif
+
+	.text
+ENTRY(__cbrtf)
+	movl	4(%esp), %eax
+	xorl	%ecx, %ecx
+	movl	%eax, %edx
+	andl	$0x7fffffff, %eax
+	jz	1f
+	cmpl	$0x7f800000, %eax
+	jae	1f
+
+#ifdef PIC
+	pushl	%ebx
+	call	3f
+3:	popl	%ebx
+	addl    $_GLOBAL_OFFSET_TABLE_+[.-3b], %ebx
+#endif
+
+	cmpl	$0x00800000, %eax
+	jae	2f
+
+#ifdef PIC
+	flds	8(%esp)
+#else
+	flds	4(%esp)
+#endif
+	fmuls	MO(two25)
+	movl	$-25, %ecx
+#ifdef PIC
+	fstps	8(%esp)
+	movl	8(%esp), %eax
+#else
+	fstps	4(%esp)
+	movl	4(%esp), %eax
+#endif
+	movl	%eax, %edx
+	andl	$0x7fffffff, %eax
+
+2:	shrl	$23, %eax
+	andl	$0x807fffff, %edx
+	subl	$126, %eax
+	orl	$0x3f000000, %edx
+	addl	%eax, %ecx
+#ifdef PIC
+	movl	%edx, 8(%esp)
+
+	flds	8(%esp)			/* xm */
+#else
+	movl	%edx, 4(%esp)
+
+	flds	4(%esp)			/* xm */
+#endif
+	fabs
+
+	/* The following code has two tracks:
+	    a) compute the normalized cbrt value
+	    b) compute xe/3 and xe%3
+	   The right track computes the value for b) and this is done
+	   in an optimized way by avoiding division.
+
+	   But why two tracks at all?  Very easy: efficiency.  Some FP
+	   instruction can overlap with a certain amount of integer (and
+	   FP) instructions.  So we get (except for the imull) all
+	   instructions for free.  */
+
+	fld	%st(0)			/* xm : xm */
+	fmull	MO(f3)			/* f3*xm : xm */
+			movl	$1431655766, %eax
+	fsubrl	MO(f2)			/* f2-f3*xm : xm */
+			imull	%ecx
+	fmul	%st(1)			/* (f2-f3*xm)*xm : xm */
+			movl	%ecx, %eax
+	faddl	MO(f1)			/* u:=f1+(f2-f3*xm)*xm : xm */
+			sarl	$31, %eax
+	fld	%st			/* u : u : xm */
+			subl	%eax, %edx
+	fmul	%st(1)			/* u*u : u : xm */
+	fld	%st(2)			/* xm : u*u : u : xm */
+	fadd	%st			/* 2*xm : u*u : u : xm */
+	fxch	%st(1)			/* u*u : 2*xm : u : xm */
+	fmul	%st(2)			/* t2:=u*u*u : 2*xm : u : xm */
+			movl	%edx, %eax
+	fadd	%st, %st(1)		/* t2 : t2+2*xm : u : xm */
+			leal	(%edx,%edx,2),%edx
+	fadd	%st(0)			/* 2*t2 : t2+2*xm : u : xm */
+			subl	%edx, %ecx
+	faddp	%st, %st(3)		/* t2+2*xm : u : 2*t2+xm */
+			shll	$3, %ecx
+	fmulp				/* u*(t2+2*xm) : 2*t2+xm */
+	fdivp	%st, %st(1)		/* u*(t2+2*xm)/(2*t2+xm) */
+	fmull	MOX(16+factor,%ecx)	/* u*(t2+2*xm)/(2*t2+xm)*FACT */
+	pushl	%eax
+	fildl	(%esp)			/* xe/3 : u*(t2+2*xm)/(2*t2+xm)*FACT */
+	fxch				/* u*(t2+2*xm)/(2*t2+xm)*FACT : xe/3 */
+	fscale				/* u*(t2+2*xm)/(2*t2+xm)*FACT*2^xe/3 */
+	popl	%edx
+#ifdef PIC
+	movl	8(%esp), %eax
+	popl	%ebx
+#else
+	movl	4(%esp), %eax
+#endif
+	testl	%eax, %eax
+	fstp	%st(1)
+	jns	4f
+	fchs
+4:	ret
+
+	/* Return the argument.  */
+1:	flds	4(%esp)
+	ret
+END(__cbrtf)
+weak_alias (__cbrtf, cbrtf)
diff --git a/sysdeps/libm-i387/s_cbrtl.S b/sysdeps/libm-i387/s_cbrtl.S
new file mode 100644
index 0000000000..6a3b9a8dc5
--- /dev/null
+++ b/sysdeps/libm-i387/s_cbrtl.S
@@ -0,0 +1,229 @@
+/* Compute cubic root of long double value.
+   Copyright (C) 1997 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Dirk Alboth <dirka@uni-paderborn.de> and
+   Ulrich Drepper <drepper@cygnus.com>, 1997.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 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
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include <machine/asm.h>
+
+#ifdef __ELF__
+        .section .rodata
+#else
+        .text
+#endif
+
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f8,@object)
+f8:	.tfloat 0.161617097923756032
+	ASM_SIZE_DIRECTIVE(f8)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f7,@object)
+f7:	.tfloat -0.988553671195413709
+	ASM_SIZE_DIRECTIVE(f7)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f6,@object)
+f6:	.tfloat 2.65298938441952296
+	ASM_SIZE_DIRECTIVE(f6)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f5,@object)
+f5:	.tfloat -4.11151425200350531
+	ASM_SIZE_DIRECTIVE(f5)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f4,@object)
+f4:	.tfloat 4.09559907378707839
+	ASM_SIZE_DIRECTIVE(f4)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f3,@object)
+f3:	.tfloat -2.82414939754975962
+	ASM_SIZE_DIRECTIVE(f3)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f2,@object)
+f2:	.tfloat 1.67595307700780102
+	ASM_SIZE_DIRECTIVE(f2)
+        .align ALIGNARG(4)
+        ASM_TYPE_DIRECTIVE(f1,@object)
+f1:	.tfloat 0.338058687610520237
+	ASM_SIZE_DIRECTIVE(f1)
+
+#define CBRT2		1.2599210498948731648
+#define ONE_CBRT2	0.793700525984099737355196796584
+#define SQR_CBRT2	1.5874010519681994748
+#define ONE_SQR_CBRT2	0.629960524947436582364439673883
+
+	/* We make the entries in the following table all 16 bytes
+	   wide to avoid having to implement a multiplication by 10.  */
+	ASM_TYPE_DIRECTIVE(factor,@object)
+        .align ALIGNARG(4)
+factor:	.tfloat ONE_SQR_CBRT2
+	.byte 0, 0, 0, 0, 0, 0
+	.tfloat ONE_CBRT2
+	.byte 0, 0, 0, 0, 0, 0
+	.tfloat 1.0
+	.byte 0, 0, 0, 0, 0, 0
+	.tfloat CBRT2
+	.byte 0, 0, 0, 0, 0, 0
+	.tfloat SQR_CBRT2
+	ASM_SIZE_DIRECTIVE(factor)
+
+        ASM_TYPE_DIRECTIVE(two64,@object)
+        .align ALIGNARG(4)
+two64:  .byte 0, 0, 0, 0, 0, 0, 0xf0, 0x43
+        ASM_SIZE_DIRECTIVE(two64)
+
+#ifdef PIC
+#define MO(op) op##@GOTOFF(%ebx)
+#define MOX(op,x) op##@GOTOFF(%ebx,x,1)
+#else
+#define MO(op) op
+#define MOX(op,x) op(x)
+#endif
+
+	.text
+ENTRY(__cbrtl)
+	movl	4(%esp), %ecx
+	movl	12(%esp), %eax
+	orl	8(%esp), %ecx
+	movl	%eax, %edx
+	andl	$0x7fff, %eax
+	orl	%eax, %ecx
+	jz	1f
+	xorl	%ecx, %ecx
+	cmpl	$0x7fff, %eax
+	je	1f
+
+#ifdef PIC
+	pushl	%ebx
+	call	3f
+3:	popl	%ebx
+	addl    $_GLOBAL_OFFSET_TABLE_+[.-3b], %ebx
+#endif
+
+	cmpl	$0, %eax
+	jne	2f
+
+#ifdef PIC
+	fldt	8(%esp)
+#else
+	fldt	4(%esp)
+#endif
+	fmull	MO(two64)
+	movl	$-64, %ecx
+#ifdef PIC
+	fstpt	8(%esp)
+	movl	16(%esp), %eax
+#else
+	fstpt	4(%esp)
+	movl	12(%esp), %eax
+#endif
+	movl	%eax, %edx
+	andl	$0x7fff, %eax
+
+2:	andl	$0x8000, %edx
+	subl	$16382, %eax
+	orl	$0x3ffe, %edx
+	addl	%eax, %ecx
+#ifdef PIC
+	movl	%edx, 16(%esp)
+
+	fldt	8(%esp)			/* xm */
+#else
+	movl	%edx, 12(%esp)
+
+	fldt	4(%esp)			/* xm */
+#endif
+	fabs
+
+	/* The following code has two tracks:
+	    a) compute the normalized cbrt value
+	    b) compute xe/3 and xe%3
+	   The right track computes the value for b) and this is done
+	   in an optimized way by avoiding division.
+
+	   But why two tracks at all?  Very easy: efficiency.  Some FP
+	   instruction can overlap with a certain amount of integer (and
+	   FP) instructions.  So we get (except for the imull) all
+	   instructions for free.  */
+
+	fldt	MO(f8)			/* f8 : xm */
+	fmul	%st(1)			/* f8*xm : xm */
+
+	fldt	MO(f7)
+	faddp				/* f7+f8*xm : xm */
+	fmul	%st(1)			/* (f7+f8*xm)*xm : xm */
+			movl	$1431655766, %eax
+	fldt	MO(f6)
+	faddp				/* f6+(f7+f8*xm)*xm : xm */
+			imull	%ecx
+	fmul	%st(1)			/* (f6+(f7+f8*xm)*xm)*xm : xm */
+			movl	%ecx, %eax
+	fldt	MO(f5)
+	faddp				/* f5+(f6+(f7+f8*xm)*xm)*xm : xm */
+			sarl	$31, %eax
+	fmul	%st(1)			/* (f5+(f6+(f7+f8*xm)*xm)*xm)*xm : xm */
+			subl	%eax, %edx
+	fldt	MO(f4)
+	faddp				/* f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm : xm */
+	fmul	%st(1)			/* (f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm)*xm : xm */
+	fldt	MO(f3)
+	faddp				/* f3+(f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm)*xm : xm */
+	fmul	%st(1)			/* (f3+(f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm)*xm)*xm : xm */
+	fldt	MO(f2)
+	faddp				/* f2+(f3+(f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm)*xm)*xm : xm */
+	fmul	%st(1)			/* (f2+(f3+(f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm)*xm)*xm)*xm : xm */
+	fldt	MO(f1)
+	faddp				/* u:=f1+(f2+(f3+(f4+(f5+(f6+(f7+f8*xm)*xm)*xm)*xm)*xm)*xm)*xm : xm */
+
+	fld	%st			/* u : u : xm */
+	fmul	%st(1)			/* u*u : u : xm */
+	fld	%st(2)			/* xm : u*u : u : xm */
+	fadd	%st			/* 2*xm : u*u : u : xm */
+	fxch	%st(1)			/* u*u : 2*xm : u : xm */
+	fmul	%st(2)			/* t2:=u*u*u : 2*xm : u : xm */
+			movl	%edx, %eax
+	fadd	%st, %st(1)		/* t2 : t2+2*xm : u : xm */
+			leal	(%edx,%edx,2),%edx
+	fadd	%st(0)			/* 2*t2 : t2+2*xm : u : xm */
+			subl	%edx, %ecx
+	faddp	%st, %st(3)		/* t2+2*xm : u : 2*t2+xm */
+			shll	$4, %ecx
+	fmulp				/* u*(t2+2*xm) : 2*t2+xm */
+	fdivp	%st, %st(1)		/* u*(t2+2*xm)/(2*t2+xm) */
+	fldt	MOX(32+factor,%ecx)
+	fmulp				/* u*(t2+2*xm)/(2*t2+xm)*FACT */
+	pushl	%eax
+	fildl	(%esp)			/* xe/3 : u*(t2+2*xm)/(2*t2+xm)*FACT */
+	fxch				/* u*(t2+2*xm)/(2*t2+xm)*FACT : xe/3 */
+	fscale				/* u*(t2+2*xm)/(2*t2+xm)*FACT*2^xe/3 */
+	popl	%edx
+#ifdef PIC
+	movl	16(%esp), %eax
+	popl	%ebx
+#else
+	movl	12(%esp), %eax
+#endif
+	testl	$0x8000, %eax
+	fstp	%st(1)
+	jz	4f
+	fchs
+4:	ret
+
+	/* Return the argument.  */
+1:	fldt	4(%esp)
+	ret
+END(__cbrtl)
+weak_alias (__cbrtl, cbrtl)