183 lines
4.5 KiB
ArmAsm
183 lines
4.5 KiB
ArmAsm
/* Compute cubic root of float value.
|
|
Copyright (C) 1997, 2005 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 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, 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
|
|
cfi_adjust_cfa_offset (4)
|
|
cfi_rel_offset (ebx, 0)
|
|
LOAD_PIC_REG (bx)
|
|
#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
|
|
cfi_adjust_cfa_offset (4)
|
|
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
|
|
cfi_adjust_cfa_offset (-4)
|
|
#ifdef PIC
|
|
movl 8(%esp), %eax
|
|
popl %ebx
|
|
cfi_adjust_cfa_offset (-4)
|
|
cfi_restore (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)
|