glibc/sysdeps/i386/fpu/e_powl.S
Joseph Myers 4da6db5188 Fix pow overflow in non-default rounding modes (bug 16315).
This patch fixes bug 16315, bad pow handling of overflow/underflow in
non-default rounding modes.  Tests of pow are duly converted to
ALL_RM_TEST to run all tests in all rounding modes.

There are two main issues here.  First, various implementations
compute a negative result by negating a positive result, but this
yields inappropriate overflow / underflow values for directed
rounding, so either overflow / underflow results need recomputing in
the correct sign, or the relevant overflowing / underflowing operation
needs to be made to have a result of the correct sign.  Second, the
dbl-64 implementation sets FE_TONEAREST internally; in the overflow /
underflow case, the result needs recomputing in the original rounding
mode.

Tested x86_64 and x86 and ulps updated accordingly.

	[BZ #16315]
	* sysdeps/i386/fpu/e_pow.S (__ieee754_pow): Ensure possibly
	overflowing or underflowing operations take place with sign of
	result.
	* sysdeps/i386/fpu/e_powf.S (__ieee754_powf): Likewise.
	* sysdeps/i386/fpu/e_powl.S (__ieee754_powl): Likewise.
	* sysdeps/ieee754/dbl-64/e_pow.c: Include <math.h>.
	(__ieee754_pow): Recompute overflowing and underflowing results in
	original rounding mode.
	* sysdeps/x86/fpu/powl_helper.c: Include <stdbool.h>.
	(__powl_helper): Allow negative argument X and scale negated value
	as needed.  Avoid passing value outside [-1, 1] to f2xm1.
	* sysdeps/x86_64/fpu/e_powl.S (__ieee754_powl): Ensure possibly
	overflowing or underflowing operations take place with sign of
	result.
	* sysdeps/x86_64/fpu/multiarch/e_pow.c [HAVE_FMA4_SUPPORT]:
	Include <math.h>.
	* math/auto-libm-test-in: Add more tests of pow.
	* math/auto-libm-test-out: Regenerated.
	* math/libm-test.inc (pow_test): Use ALL_RM_TEST.
	(pow_tonearest_test_data): Remove.
	(pow_test_tonearest): Likewise.
	(pow_towardzero_test_data): Likewise.
	(pow_test_towardzero): Likewise.
	(pow_downward_test_data): Likewise.
	(pow_test_downward): Likewise.
	(pow_upward_test_data): Likewise.
	(pow_test_upward): Likewise.
	(main): Don't call removed functions.
	* sysdeps/i386/fpu/libm-test-ulps: Update.
	* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
2014-06-23 20:12:33 +00:00

435 lines
8.8 KiB
ArmAsm
Raw Blame History

/* ix87 specific implementation of pow function.
Copyright (C) 1996-2014 Free Software Foundation, Inc.
This file is part of the GNU C Library.
Contributed by Ulrich Drepper <drepper@cygnus.com>, 1996.
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, see
<http://www.gnu.org/licenses/>. */
#include <machine/asm.h>
.section .rodata.cst8,"aM",@progbits,8
.p2align 3
.type one,@object
one: .double 1.0
ASM_SIZE_DIRECTIVE(one)
.type p3,@object
p3: .byte 0, 0, 0, 0, 0, 0, 0x20, 0x40
ASM_SIZE_DIRECTIVE(p3)
.type p63,@object
p63: .byte 0, 0, 0, 0, 0, 0, 0xe0, 0x43
ASM_SIZE_DIRECTIVE(p63)
.type p64,@object
p64: .byte 0, 0, 0, 0, 0, 0, 0xf0, 0x43
ASM_SIZE_DIRECTIVE(p64)
.type p78,@object
p78: .byte 0, 0, 0, 0, 0, 0, 0xd0, 0x44
ASM_SIZE_DIRECTIVE(p78)
.type pm79,@object
pm79: .byte 0, 0, 0, 0, 0, 0, 0, 0x3b
ASM_SIZE_DIRECTIVE(pm79)
.section .rodata.cst16,"aM",@progbits,16
.p2align 3
.type infinity,@object
inf_zero:
infinity:
.byte 0, 0, 0, 0, 0, 0, 0xf0, 0x7f
ASM_SIZE_DIRECTIVE(infinity)
.type zero,@object
zero: .double 0.0
ASM_SIZE_DIRECTIVE(zero)
.type minf_mzero,@object
minf_mzero:
minfinity:
.byte 0, 0, 0, 0, 0, 0, 0xf0, 0xff
mzero:
.byte 0, 0, 0, 0, 0, 0, 0, 0x80
ASM_SIZE_DIRECTIVE(minf_mzero)
#ifdef PIC
# define MO(op) op##@GOTOFF(%ecx)
# define MOX(op,x,f) op##@GOTOFF(%ecx,x,f)
#else
# define MO(op) op
# define MOX(op,x,f) op(,x,f)
#endif
.text
ENTRY(__ieee754_powl)
fldt 16(%esp) // y
fxam
#ifdef PIC
LOAD_PIC_REG (cx)
#endif
fnstsw
movb %ah, %dl
andb $0x45, %ah
cmpb $0x40, %ah // is y == 0 ?
je 11f
cmpb $0x05, %ah // is y == <EFBFBD>inf ?
je 12f
cmpb $0x01, %ah // is y == NaN ?
je 30f
fldt 4(%esp) // x : y
subl $8,%esp
cfi_adjust_cfa_offset (8)
fxam
fnstsw
movb %ah, %dh
andb $0x45, %ah
cmpb $0x40, %ah
je 20f // x is <EFBFBD>0
cmpb $0x05, %ah
je 15f // x is <EFBFBD>inf
cmpb $0x01, %ah
je 32f // x is NaN
fxch // y : x
/* fistpll raises invalid exception for |y| >= 1L<<63. */
fld %st // y : y : x
fabs // |y| : y : x
fcompl MO(p63) // y : x
fnstsw
sahf
jnc 2f
/* First see whether `y' is a natural number. In this case we
can use a more precise algorithm. */
fld %st // y : y : x
fistpll (%esp) // y : x
fildll (%esp) // int(y) : y : x
fucomp %st(1) // y : x
fnstsw
sahf
je 9f
// If y has absolute value at most 0x1p-79, then any finite
// nonzero x will result in 1. Saturate y to those bounds to
// avoid underflow in the calculation of y*log2(x).
fld %st // y : y : x
fabs // |y| : y : x
fcompl MO(pm79) // y : x
fnstsw
sahf
jnc 3f
fstp %st(0) // pop y
fldl MO(pm79) // 0x1p-79 : x
testb $2, %dl
jnz 3f // y > 0
fchs // -0x1p-79 : x
jmp 3f
9: /* OK, we have an integer value for y. Unless very small
(we use < 8), use the algorithm for real exponent to avoid
accumulation of errors. */
fld %st // y : y : x
fabs // |y| : y : x
fcompl MO(p3) // y : x
fnstsw
sahf
jnc 3f
popl %eax
cfi_adjust_cfa_offset (-4)
popl %edx
cfi_adjust_cfa_offset (-4)
orl $0, %edx
fstp %st(0) // x
jns 4f // y >= 0, jump
fdivrl MO(one) // 1/x (now referred to as x)
negl %eax
adcl $0, %edx
negl %edx
4: fldl MO(one) // 1 : x
fxch
/* If y is even, take the absolute value of x. Otherwise,
ensure all intermediate values that might overflow have the
sign of x. */
testb $1, %al
jnz 6f
fabs
6: shrdl $1, %edx, %eax
jnc 5f
fxch
fabs
fmul %st(1) // x : ST*x
fxch
5: fld %st // x : x : ST*x
fabs // |x| : x : ST*x
fmulp // |x|*x : ST*x
shrl $1, %edx
movl %eax, %ecx
orl %edx, %ecx
jnz 6b
fstp %st(0) // ST*x
ret
/* y is <20>NAN */
30: fldt 4(%esp) // x : y
fldl MO(one) // 1.0 : x : y
fucomp %st(1) // x : y
fnstsw
sahf
je 31f
fxch // y : x
31: fstp %st(1)
ret
cfi_adjust_cfa_offset (8)
32: addl $8, %esp
cfi_adjust_cfa_offset (-8)
fstp %st(1)
ret
cfi_adjust_cfa_offset (8)
.align ALIGNARG(4)
2: // y is a large integer (absolute value at least 1L<<63).
// If y has absolute value at least 1L<<78, then any finite
// nonzero x will result in 0 (underflow), 1 or infinity (overflow).
// Saturate y to those bounds to avoid overflow in the calculation
// of y*log2(x).
fld %st // y : y : x
fabs // |y| : y : x
fcompl MO(p78) // y : x
fnstsw
sahf
jc 3f
fstp %st(0) // pop y
fldl MO(p78) // 1L<<78 : x
testb $2, %dl
jz 3f // y > 0
fchs // -(1L<<78) : x
.align ALIGNARG(4)
3: /* y is a real number. */
subl $28, %esp
cfi_adjust_cfa_offset (28)
fstpt 12(%esp) // x
fstpt (%esp) // <empty>
call HIDDEN_JUMPTARGET (__powl_helper) // <result>
addl $36, %esp
cfi_adjust_cfa_offset (-36)
ret
// pow(x,<EFBFBD>0) = 1
.align ALIGNARG(4)
11: fstp %st(0) // pop y
fldl MO(one)
ret
// y == <EFBFBD>inf
.align ALIGNARG(4)
12: fstp %st(0) // pop y
fldl MO(one) // 1
fldt 4(%esp) // x : 1
fabs // abs(x) : 1
fucompp // < 1, == 1, or > 1
fnstsw
andb $0x45, %ah
cmpb $0x45, %ah
je 13f // jump if x is NaN
cmpb $0x40, %ah
je 14f // jump if |x| == 1
shlb $1, %ah
xorb %ah, %dl
andl $2, %edx
fldl MOX(inf_zero, %edx, 4)
ret
.align ALIGNARG(4)
14: fldl MO(one)
ret
.align ALIGNARG(4)
13: fldt 4(%esp) // load x == NaN
ret
cfi_adjust_cfa_offset (8)
.align ALIGNARG(4)
// x is <EFBFBD>inf
15: fstp %st(0) // y
testb $2, %dh
jz 16f // jump if x == +inf
// fistpll raises invalid exception for |y| >= 1L<<63, but y
// may be odd unless we know |y| >= 1L<<64.
fld %st // y : y
fabs // |y| : y
fcompl MO(p64) // y
fnstsw
sahf
jnc 16f
fldl MO(p63) // p63 : y
fxch // y : p63
fprem // y%p63 : p63
fstp %st(1) // y%p63
// We must find out whether y is an odd integer.
fld %st // y : y
fistpll (%esp) // y
fildll (%esp) // int(y) : y
fucompp // <empty>
fnstsw
sahf
jne 17f
// OK, the value is an integer, but is it odd?
popl %eax
cfi_adjust_cfa_offset (-4)
popl %edx
cfi_adjust_cfa_offset (-4)
andb $1, %al
jz 18f // jump if not odd
// It's an odd integer.
shrl $31, %edx
fldl MOX(minf_mzero, %edx, 8)
ret
cfi_adjust_cfa_offset (8)
.align ALIGNARG(4)
16: fcompl MO(zero)
addl $8, %esp
cfi_adjust_cfa_offset (-8)
fnstsw
shrl $5, %eax
andl $8, %eax
fldl MOX(inf_zero, %eax, 1)
ret
cfi_adjust_cfa_offset (8)
.align ALIGNARG(4)
17: shll $30, %edx // sign bit for y in right position
addl $8, %esp
cfi_adjust_cfa_offset (-8)
18: shrl $31, %edx
fldl MOX(inf_zero, %edx, 8)
ret
cfi_adjust_cfa_offset (8)
.align ALIGNARG(4)
// x is <EFBFBD>0
20: fstp %st(0) // y
testb $2, %dl
jz 21f // y > 0
// x is <EFBFBD>0 and y is < 0. We must find out whether y is an odd integer.
testb $2, %dh
jz 25f
// fistpll raises invalid exception for |y| >= 1L<<63, but y
// may be odd unless we know |y| >= 1L<<64.
fld %st // y : y
fabs // |y| : y
fcompl MO(p64) // y
fnstsw
sahf
jnc 25f
fldl MO(p63) // p63 : y
fxch // y : p63
fprem // y%p63 : p63
fstp %st(1) // y%p63
fld %st // y : y
fistpll (%esp) // y
fildll (%esp) // int(y) : y
fucompp // <empty>
fnstsw
sahf
jne 26f
// OK, the value is an integer, but is it odd?
popl %eax
cfi_adjust_cfa_offset (-4)
popl %edx
cfi_adjust_cfa_offset (-4)
andb $1, %al
jz 27f // jump if not odd
// It's an odd integer.
// Raise divide-by-zero exception and get minus infinity value.
fldl MO(one)
fdivl MO(zero)
fchs
ret
cfi_adjust_cfa_offset (8)
25: fstp %st(0)
26: addl $8, %esp
cfi_adjust_cfa_offset (-8)
27: // Raise divide-by-zero exception and get infinity value.
fldl MO(one)
fdivl MO(zero)
ret
cfi_adjust_cfa_offset (8)
.align ALIGNARG(4)
// x is <EFBFBD>0 and y is > 0. We must find out whether y is an odd integer.
21: testb $2, %dh
jz 22f
// fistpll raises invalid exception for |y| >= 1L<<63, but y
// may be odd unless we know |y| >= 1L<<64.
fld %st // y : y
fcompl MO(p64) // y
fnstsw
sahf
jnc 22f
fldl MO(p63) // p63 : y
fxch // y : p63
fprem // y%p63 : p63
fstp %st(1) // y%p63
fld %st // y : y
fistpll (%esp) // y
fildll (%esp) // int(y) : y
fucompp // <empty>
fnstsw
sahf
jne 23f
// OK, the value is an integer, but is it odd?
popl %eax
cfi_adjust_cfa_offset (-4)
popl %edx
cfi_adjust_cfa_offset (-4)
andb $1, %al
jz 24f // jump if not odd
// It's an odd integer.
fldl MO(mzero)
ret
cfi_adjust_cfa_offset (8)
22: fstp %st(0)
23: addl $8, %esp // Don't use 2 x pop
cfi_adjust_cfa_offset (-8)
24: fldl MO(zero)
ret
END(__ieee754_powl)
strong_alias (__ieee754_powl, __powl_finite)