powerpc fenv

powerpc fenv

[Szabolcs Nagy]

[-- Attachment #1: Type: text/plain, Size: 713 bytes --]

i wrote an initial version of fenv for ppc
(absolutely not tested)

i followed glibc and used double for fenv_t
(it makes fesetenv and fegetenv a bit shorter
than with using a 32bit type)

i don't know what section the target of the
FE_DFL_ENV pointer should go and what's the
proper asm syntax for objects
(i put a 0 value in .sdata for now)

ps. the current feholdexcept is wrong, it
should feclearexcept(FE_ALL_EXCEPT)

pps. i put a FENV_ACCESS ON pragma in feupdateenv.
it's silly, since there is no fp operation that
could go wrong there, but c99 implies that every
time c code may access the fenv it must be present,
so probably the other c functions need it as well
(it does not hurt as it's a nop on gcc)

[-- Attachment #2: fenv.h --]
[-- Type: text/x-chdr, Size: 807 bytes --]

#define FE_TONEAREST	0
#define FE_TOWARDZERO	1
#define FE_UPWARD	2
#define FE_DOWNWARD	3

#define FE_INEXACT	0x02000000
#define FE_DIVBYZERO	0x04000000
#define FE_UNDERFLOW	0x08000000
#define FE_OVERFLOW	0x10000000
#define FE_INVALID	0x20000000

#define FE_ALL_EXCEPT	0x3e000000

#ifdef _GNU_SOURCE
#define FE_INVALID_SNAN		0x01000000
#define FE_INVALID_ISI		0x00800000
#define FE_INVALID_IDI		0x00400000
#define FE_INVALID_ZDZ		0x00200000
#define FE_INVALID_IMZ		0x00100000
#define FE_INVALID_COMPARE	0x00080000
#define FE_INVALID_SOFTWARE	0x00000400
#define FE_INVALID_SQRT		0x00000200
#define FE_INVALID_INTEGER_CONVERSION	0x00000100

#define FE_ALL_INVALID		0x01f80700
#endif

typedef unsigned int fexcept_t;
typedef double fenv_t;

extern const fenv_t __fe_dfl_env;
#define FE_DFL_ENV (&__fe_dfl_env)


[-- Attachment #3: fenv.s --]
[-- Type: text/plain, Size: 1720 bytes --]

.global feclearexcept
.type feclearexcept,@function
feclearexcept:
	# if (r3 & FE_INVALID) r3 |= all_invalid_flags
	andis. 0,3,0x2000
	stwu 1,-16(1)
	beq- 0,1f
	oris 3,3,0x01f8
	ori  3,3,0x0700
1:
	# note: fpscr contains various fpu status and control
	# flags and we dont check if r3 may alter other flags
	# than the exception related ones
	# fpscr &= ~r3
	mffs 0
	stfd 0,8(1)
	lwz 9,12(1)
	andc 9,9,3
	stw 9,12(1)
	lfd 0,8(1)
	mtfsf 255,0

	# return 0
	li 3,0
	addi 1,1,16
	blr

.global feraiseexcept
.type feraiseexcept,@function
feraiseexcept:
	# if (r3 & FE_INVALID) r3 |= software_invalid_flag
	andis. 0,3,0x2000
	stwu 1,-16(1)
	beq- 0,1f
	ori 3,3,0x0400
1:
	# fpscr |= r3
	mffs 0
	stfd 0,8(1)
	lwz 9,12(1)
	or 9,9,3
	stw 9,12(1)
	lfd 0,8(1)
	mtfsf 255,0

	# return 0
	li 3,0
	addi 1,1,16
	blr

.global fetestexcept
.type fetestexcept,@function
fetestexcept:
	# return r3 & fpscr
	stwu 1,-16(1)
	mffs 0
	stfd 0,8(1)
	lwz 9,12(1)
	addi 1,1,16
	and 3,3,9
	blr

.global fegetround
.type fegetround,@function
fegetround:
	# return fpscr & 3
	stwu 1,-16(1)
	mffs 0
	stfd 0,8(1)
	lwz 3,12(1)
	addi 1,1,16
	clrlwi 3,3,30
	blr

.global fesetround
.type fesetround,@function
fesetround:
	# note: invalid input is not checked, r3 < 4 must hold
	# fpscr = (fpscr & -4U) | r3
	stwu 1,-16(1)
	mffs 0
	stfd 0,8(1)
	lwz 9,12(1)
	clrrwi 9,9,2
	or 9,9,3
	stw 9,12(1)
	lfd 0,8(1)
	mtfsf 255,0

	# return 0
	li 3,0
	addi 1,1,16
	blr

.global fegetenv
.type fegetenv,@function
fegetenv:
	# *r3 = fpscr
	mffs 0
	stfd 0,0(3)
	# return 0
	li 3,0
	blr

.global fesetenv
.type fesetenv,@function
fesetenv:
	# fpscr = *r3
	lfd 0,0(3)
	mtfsf 255,0
	# return 0
	li 3,0
	blr

.section .sdata
.type __fe_dfl_env,@object
__fe_dfl_env:
	.zero 8