1
0
kolibrios/programs/develop/open watcom/trunk/clib/fpu/chipr32.asm

844 lines
32 KiB
NASM
Raw Normal View History

;*****************************************************************************
;*
;* Open Watcom Project
;*
;* Portions Copyright (c) 1983-2002 Sybase, Inc. All Rights Reserved.
;*
;* ========================================================================
;*
;* This file contains Original Code and/or Modifications of Original
;* Code as defined in and that are subject to the Sybase Open Watcom
;* Public License version 1.0 (the 'License'). You may not use this file
;* except in compliance with the License. BY USING THIS FILE YOU AGREE TO
;* ALL TERMS AND CONDITIONS OF THE LICENSE. A copy of the License is
;* provided with the Original Code and Modifications, and is also
;* available at www.sybase.com/developer/opensource.
;*
;* The Original Code and all software distributed under the License are
;* distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
;* EXPRESS OR IMPLIED, AND SYBASE AND ALL CONTRIBUTORS HEREBY DISCLAIM
;* ALL SUCH WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF
;* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR
;* NON-INFRINGEMENT. Please see the License for the specific language
;* governing rights and limitations under the License.
;*
;* ========================================================================
;*
;* Description: WHEN YOU FIGURE OUT WHAT THIS FILE DOES, PLEASE
;* DESCRIBE IT HERE!
;*
;*****************************************************************************
; static char sccs_id[] = "@(#)fprem32.asm 1.5 12/22/94 12:48:07";
;
; This code is being published by Intel to users of the Pentium(tm)
; processor. Recipients are authorized to copy, modify, compile, use and
; distribute the code.
;
; Intel makes no warranty of any kind with regard to this code, including
; but not limited to, implied warranties or merchantability and fitness for
; a particular purpose. Intel assumes no responsibility for any errors that
; may appear in this code.
;
; No patent licenses are granted, express or implied.
;
;
include mdef.inc
.386
.387
;
; PRELIMINARY VERSION of the software patch for the floating
; point remainder.
;
CHECKSW MACRO
ifdef DEBUG
fnstsw [fpsw]
fnstcw [fpcw]
endif
ENDM
_DATA SEGMENT DWORD USE32 PUBLIC 'DATA'
;
; Stack variables for remainder routines.
;
FLT_SIZE EQU 12
DENOM EQU 0
DENOM_SAVE EQU DENOM + FLT_SIZE
NUMER EQU DENOM_SAVE + FLT_SIZE
PREV_CW EQU NUMER + FLT_SIZE
PATCH_CW EQU PREV_CW + 4
FPREM_SW EQU PATCH_CW + 4
STACK_SIZE EQU FPREM_SW + 4
RET_SIZE EQU 4
PUSH_SIZE EQU 4
MAIN_FUDGE EQU RET_SIZE + PUSH_SIZE + PUSH_SIZE + PUSH_SIZE
MAIN_DENOM EQU DENOM + MAIN_FUDGE
MAIN_DENOM_SAVE EQU DENOM_SAVE + MAIN_FUDGE
MAIN_NUMER EQU NUMER + MAIN_FUDGE
MAIN_PREV_CW EQU PREV_CW + MAIN_FUDGE
MAIN_PATCH_CW EQU PATCH_CW + MAIN_FUDGE
MAIN_FPREM_SW EQU FPREM_SW + MAIN_FUDGE
ONESMASK EQU 700h
fprem_risc_table DB 0, 1, 0, 0, 4, 0, 0, 7, 0, 0, 10, 0, 0, 13, 0, 0
fprem_scale DB 0, 0, 0, 0, 0, 0, 0eeh, 03fh
one_shl_64 DB 0, 0, 0, 0, 0, 0, 0f0h, 043h
one_shr_64 DB 0, 0, 0, 0, 0, 0, 0f0h, 03bh
one DB 0, 0, 0, 0, 0, 0, 0f0h, 03fh
half DB 0, 0, 0, 0, 0, 0, 0e0h, 03fh
big_number DB 0, 0, 0, 0, 0, 0, 0ffh, 0ffh, 0feh, 07fh
ifdef DEBUG
public fpcw
public fpsw
fpcw dw 0
fpsw dw 0
endif
FPU_STATE STRUC
CONTROL_WORD DW ?
reserved_1 DW ?
STATUS_WORD DD ?
TAG_WORD DW ?
reserved_3 DW ?
IP_OFFSET DD ?
CS_SLCT DW ?
OPCODE DW ?
DATA_OFFSET DD ?
OPERAND_SLCT DW ?
reserved_4 DW ?
FPU_STATE ENDS
ENV_SIZE EQU 28
_DATA ENDS
_TEXT SEGMENT DWORD USE32 PUBLIC 'CODE'
_TEXT ENDS
DGROUP GROUP _DATA
_TEXT SEGMENT DWORD USE32 PUBLIC 'CODE'
assume cs:_TEXT, ds:DGROUP, es:DGROUP, ss:nothing
fprem_common PROC NEAR
push eax
push ebx
push ecx
mov eax, [MAIN_DENOM+6+esp] ; exponent and high 16 bits of mantissa
xor eax, ONESMASK ; invert bits that have to be one
test eax, ONESMASK ; check bits that have to be one
jnz remainder_hardware_ok
shr eax, 11
and eax, 0fh
cmp byte ptr fprem_risc_table[eax], 0 ; check for (1,4,7,a,d)
jz remainder_hardware_ok
; The denominator has the bit pattern. Weed out the funny cases like NaNs
; before applying the software version. Our caller guarantees that the
; denominator is not a denormal. Here we check for:
; denominator inf, NaN, unnormal
; numerator inf, NaN, unnormal, denormal
mov eax, [MAIN_DENOM+6+esp] ; exponent and high 16 bits of mantissa
and eax, 07fff0000h ; mask the exponent only
cmp eax, 07fff0000h ; check for INF or NaN
je remainder_hardware_ok
mov eax, [MAIN_NUMER+6+esp] ; exponent and high 16 bits of mantissa
and eax, 07fff0000h ; mask the exponent only
jz remainder_hardware_ok ; jif numerator denormal
cmp eax, 07fff0000h ; check for INF or NaN
je remainder_hardware_ok
mov eax, [esp + MAIN_NUMER + 4] ; high mantissa bits - numerator
add eax, eax ; set carry if explicit bit set
jnz remainder_hardware_ok ; jmp if numerator is unnormal
mov eax, [esp + MAIN_DENOM + 4] ; high mantissa bits - denominator
add eax, eax ; set carry if explicit bit set
jnz remainder_hardware_ok ; jmp if denominator is unnormal
rem_patch:
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
and eax, 07fffh ; clear sy
add eax, 63 ; evaluate ey + 63
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
sub ebx, eax ; evaluate the exponent difference (ex - ey)
ja rem_large ; if ex > ey + 63, case of large arguments
rem_patch_loop:
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
and eax, 07fffh ; clear sy
add eax, 10 ; evaluate ey + 10
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
sub ebx, eax ; evaluate the exponent difference (ex - ey)
js remainder_hardware_ok ; safe if ey + 10 > ex
fld tbyte ptr [MAIN_NUMER+esp] ; load the numerator
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
mov ecx, ebx
sub ebx, eax
and ebx, 07h
or ebx, 04h
sub ecx, ebx
mov ebx, eax
and ebx, 08000h ; keep sy
or ecx, ebx ; merge the sign of y
mov dword ptr [MAIN_DENOM+8+esp], ecx
fld tbyte ptr [MAIN_DENOM+esp] ; load the shifted denominator
mov dword ptr [MAIN_DENOM+8+esp], eax ; restore the initial denominator
fxch
fprem ; this rem is safe
fstp tbyte ptr [MAIN_NUMER+esp] ; update the numerator
fstp st(0) ; pop the stack
jmp rem_patch_loop
rem_large:
test edx, 02h ; is denominator already saved
jnz already_saved
fld tbyte ptr[esp + MAIN_DENOM]
fstp tbyte ptr[esp + MAIN_DENOM_SAVE] ; save denominator
already_saved:
; Save user's precision control and institute 80. The fp ops in
; rem_large_loop must not round to user's precision (if it is less
; than 80) because the hardware would not have done so. We are
; aping the hardware here, which is all extended.
fnstcw [esp+MAIN_PREV_CW] ; save caller's control word
mov eax, dword ptr[esp + MAIN_PREV_CW]
or eax, 033fh ; mask exceptions, pc=80
mov [esp + MAIN_PATCH_CW], eax
fldcw [esp + MAIN_PATCH_CW]
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
and eax, 07fffh ; clear sy
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
sub ebx, eax ; evaluate the exponent difference
and ebx, 03fh
or ebx, 020h
add ebx, 1
mov ecx, ebx
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
and eax, 08000h ; keep sy
or ebx, eax ; merge the sign of y
mov dword ptr[MAIN_DENOM+8+esp], ebx ; make ey equal to ex (scaled denominator)
fld tbyte ptr [MAIN_DENOM+esp] ; load the scaled denominator
fabs
fld tbyte ptr [MAIN_NUMER+esp] ; load the numerator
fabs
rem_large_loop:
fcom
fstsw ax
and eax, 00100h
jnz rem_no_sub
fsub st, st(1)
rem_no_sub:
fxch
fmul qword ptr half
fxch
sub ecx, 1 ; decrement the loop counter
jnz rem_large_loop
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
fstp tbyte ptr[esp + MAIN_NUMER] ; save result
fstp st ; toss modified denom
fld tbyte ptr[esp + MAIN_DENOM_SAVE]
fld tbyte ptr[big_number] ; force C2 to be set
fprem
fstp st
fld tbyte ptr[esp + MAIN_NUMER] ; restore saved result
fldcw [esp + MAIN_PREV_CW] ; restore caller's control word
and ebx, 08000h ; keep sx
jz rem_done
fchs
jmp rem_done
remainder_hardware_ok:
fld tbyte ptr [MAIN_DENOM+esp] ; load the denominator
fld tbyte ptr [MAIN_NUMER+esp] ; load the numerator
fprem ; and finally do a remainder
; prem_main_routine end
rem_done:
test edx, 03h
jz rem_exit
fnstsw [esp + MAIN_FPREM_SW] ; save Q0 Q1 and Q2
test edx, 01h
jz do_not_de_scale
; De-scale the result. Go to pc=80 to prevent from fmul
; from user precision (fprem does not round the result).
fnstcw [esp + MAIN_PREV_CW] ; save callers control word
mov eax, [esp + MAIN_PREV_CW]
or eax, 0300h ; pc = 80
mov [esp + MAIN_PATCH_CW], eax
fldcw [esp + MAIN_PATCH_CW]
fmul qword ptr one_shr_64
fldcw [esp + MAIN_PREV_CW] ; restore callers CW
do_not_de_scale:
mov eax, [esp + MAIN_FPREM_SW]
fxch
fstp st
fld tbyte ptr[esp + MAIN_DENOM_SAVE]
fxch
and eax, 04300h ; restore saved Q0, Q1, Q2
sub esp, ENV_SIZE
fnstenv [esp]
and [esp].STATUS_WORD, 0bcffh
or [esp].STATUS_WORD, eax
fldenv [esp]
add esp, ENV_SIZE
rem_exit:
pop ecx
pop ebx
pop eax
CHECKSW ; debug only: save status
ret
fprem_common ENDP
comment ~****************************************************************
;
; float frem_chk (float numer, float denom)
;
public frem_chk
frem_chk PROC NEAR
push edx
sub esp, STACK_SIZE
fld dword ptr [STACK_SIZE+8+esp]
fstp tbyte ptr [NUMER+esp]
fld dword ptr [STACK_SIZE+12+esp]
fstp tbyte ptr [DENOM+esp]
mov edx, 0 ; dx = 1 if denormal extended divisor
call fprem_common
fxch
fstp st
add esp, STACK_SIZE
pop edx
ret
frem_chk ENDP
; end frem_chk
;
; double drem_chk (double numer, double denom)
;
public drem_chk
drem_chk PROC NEAR
push edx
sub esp, STACK_SIZE
fld qword ptr [STACK_SIZE+8+esp]
fstp tbyte ptr [NUMER+esp]
fld qword ptr [STACK_SIZE+16+esp]
fstp tbyte ptr [DENOM+esp]
mov edx, 0 ; dx = 1 if denormal extended divisor
call fprem_common
fxch
fstp st
add esp, STACK_SIZE
pop edx
ret
drem_chk ENDP
; end drem_chk
;
; long double lrem_chk(long double number,long double denom)
;
public lrem_chk
lrem_chk PROC NEAR
fld tbyte ptr [20+esp]
fld tbyte ptr [4+esp]
call fprem_chk
fxch
fstp st
ret
lrem_chk ENDP
**********************************************************************~
;
; FPREM: ST = remainder(ST, ST(1))
;
; Compiler version of the FPREM must preserve the arguments in the floating
; point stack.
public __fprem_chk
defpe __fprem_chk
push edx
sub esp, STACK_SIZE
fstp tbyte ptr [NUMER+esp]
fstp tbyte ptr [DENOM+esp]
xor edx, edx
; prem_main_routine begin
mov eax,[DENOM+6+esp] ; exponent and high 16 bits of mantissa
test eax,07fff0000h ; check for denormal
jz denormal
call fprem_common
add esp, STACK_SIZE
pop edx
ret
denormal:
fld tbyte ptr [DENOM+esp] ; load the denominator
fld tbyte ptr [NUMER+esp] ; load the numerator
mov eax, [DENOM+esp] ; test for whole mantissa == 0
or eax, [DENOM+4+esp] ; test for whole mantissa == 0
jz remainder_hardware_ok_l ; denominator is zero
fxch
fstp tbyte ptr[esp + DENOM_SAVE] ; save org denominator
fld tbyte ptr[esp + DENOM]
fxch
or edx, 02h
;
; For this we need pc=80. Also, mask exceptions so we don't take any
; denormal operand exceptions. It is guaranteed that the descaling
; later on will take underflow, which is what the hardware would have done
; on a normal fprem.
;
fnstcw [PREV_CW+esp] ; save caller's control word
mov eax, [PREV_CW+esp]
or eax, 0033fh ; mask exceptions, pc=80
mov [PATCH_CW+esp], eax
fldcw [PATCH_CW+esp] ; mask exceptions & pc=80
; The denominator is a denormal. For most numerators, scale both numerator
; and denominator to get rid of denormals. Then execute the common code
; with the flag set to indicate that the result must be de-scaled.
; For large numerators this won't work because the scaling would cause
; overflow. In this case we know the numerator is large, the denominator
; is small (denormal), so the exponent difference is also large. This means
; the rem_large code will be used and this code depends on the difference
; in exponents modulo 64. Adding 64 to the denominators exponent
; doesn't change the modulo 64 difference. So we can scale the denominator
; by 64, making it not denormal, and this won't effect the result.
;
; To start with, figure out if numerator is large
mov eax, [esp + NUMER + 8] ; load numerator exponent
and eax, 7fffh ; isolate numerator exponent
cmp eax, 7fbeh ; compare Nexp to Maxexp-64
ja big_numer_rem_de ; jif big numerator
; So the numerator is not large scale both numerator and denominator
or edx, 1 ; edx = 1, if denormal extended divisor
fmul qword ptr one_shl_64 ; make numerator not denormal
fstp tbyte ptr[esp + NUMER]
fmul qword ptr one_shl_64 ; make denominator not denormal
fstp tbyte ptr[esp + DENOM]
jmp scaling_done
; The numerator is large. Scale only the denominator, which will not
; change the result which we know will be partial. Set the scale flag
; to false.
big_numer_rem_de:
; We must do this with pc=80 to avoid rounding to single/double.
; In this case we do not mask exceptions so that we will take
; denormal operand, as would the hardware.
fnstcw [PREV_CW+esp] ; save caller's control word
mov eax, [PREV_CW+esp]
or eax, 00300h ; pc=80
mov [PATCH_CW+esp], eax
fldcw [PATCH_CW+esp] ; pc=80
fstp st ; Toss numerator
fmul qword ptr one_shl_64 ; make denominator not denormal
fstp tbyte ptr[esp + DENOM]
; Restore the control word which was fiddled to scale at 80-bit precision.
; Then call the common code.
scaling_done:
fldcw [esp + PREV_CW] ; restore callers control word
call fprem_common
add esp, STACK_SIZE
pop edx
ret
remainder_hardware_ok_l:
fprem ; and finally do a remainder
CHECKSW
add esp, STACK_SIZE
pop edx
ret
__fprem_chk ENDP
; end fprem_chk
;
; FPREM1 code begins here
;
fprem1_common PROC NEAR
push eax
push ebx
push ecx
mov eax, [MAIN_DENOM+6+esp] ; exponent and high 16 bits of mantissa
xor eax, ONESMASK ; invert bits that have to be one
test eax, ONESMASK ; check bits that have to be one
jnz remainder1_hardware_ok
shr eax, 11
and eax, 0fh
cmp byte ptr fprem_risc_table[eax], 0 ; check for (1,4,7,a,d)
jz remainder1_hardware_ok
; The denominator has the bit pattern. Weed out the funny cases like NaNs
; before applying the software version. Our caller guarantees that the
; denominator is not a denormal. Here we check for:
; denominator inf, NaN, unnormal
; numerator inf, NaN, unnormal, denormal
mov eax, [MAIN_DENOM+6+esp] ; exponent and high 16 bits of mantissa
and eax, 07fff0000h ; mask the exponent only
cmp eax, 07fff0000h ; check for INF or NaN
je remainder1_hardware_ok
mov eax, [MAIN_NUMER+6+esp] ; exponent and high 16 bits of mantissa
and eax, 07fff0000h ; mask the exponent only
jz remainder1_hardware_ok ; jif numerator denormal
cmp eax, 07fff0000h ; check for INF or NaN
je remainder1_hardware_ok
mov eax, [esp + MAIN_NUMER + 4] ; high mantissa bits - numerator
add eax, eax ; set carry if explicit bit set
jnz remainder1_hardware_ok ; jmp if numerator is unnormal
mov eax, [esp + MAIN_DENOM + 4] ; high mantissa bits - denominator
add eax, eax ; set carry if explicit bit set
jnz remainder1_hardware_ok ; jmp if denominator is unnormal
rem1_patch:
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
and eax, 07fffh ; clear sy
add eax, 63 ; evaluate ey + 63
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
sub ebx, eax ; evaluate the exponent difference (ex - ey)
ja rem1_large ; if ex > ey + 63, case of large arguments
rem1_patch_loop:
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
and eax, 07fffh ; clear sy
add eax, 10 ; evaluate ey + 10
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
sub ebx, eax ; evaluate the exponent difference (ex - ey)
js remainder1_hardware_ok ; safe if ey + 10 > ex
fld tbyte ptr [MAIN_NUMER+esp] ; load the numerator
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
mov ecx, ebx
sub ebx, eax
and ebx, 07h
or ebx, 04h
sub ecx, ebx
mov ebx, eax
and ebx, 08000h ; keep sy
or ecx, ebx ; merge the sign of y
mov dword ptr [MAIN_DENOM+8+esp], ecx
fld tbyte ptr [MAIN_DENOM+esp] ; load the shifted denominator
mov dword ptr [MAIN_DENOM+8+esp], eax ; restore the initial denominator
fxch
fprem ; this rem is safe
fstp tbyte ptr [MAIN_NUMER+esp] ; update the numerator
fstp st(0) ; pop the stack
jmp rem1_patch_loop
rem1_large:
test ebx, 02h ; is denominator already saved
jnz already_saved1
fld tbyte ptr[esp + MAIN_DENOM]
fstp tbyte ptr[esp + MAIN_DENOM_SAVE] ; save denominator
already_saved1:
; Save user's precision control and institute 80. The fp ops in
; rem1_large_loop must not round to user's precision (if it is less
; than 80) because the hardware would not have done so. We are
; aping the hardware here, which is all extended.
fnstcw [esp+MAIN_PREV_CW] ; save caller's control word
mov eax, dword ptr[esp + MAIN_PREV_CW]
or eax, 033fh ; mask exceptions, pc=80
mov [esp + MAIN_PATCH_CW], eax
fldcw [esp + MAIN_PATCH_CW]
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
and eax, 07fffh ; clear sy
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
sub ebx, eax ; evaluate the exponent difference
and ebx, 03fh
or ebx, 020h
add ebx, 1
mov ecx, ebx
mov eax, [MAIN_DENOM+8+esp] ; sign and exponent of y (denominator)
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
and ebx, 07fffh ; clear sx
and eax, 08000h ; keep sy
or ebx, eax ; merge the sign of y
mov dword ptr[MAIN_DENOM+8+esp], ebx ; make ey equal to ex (scaled denominator)
fld tbyte ptr [MAIN_DENOM+esp] ; load the scaled denominator
fabs
fld tbyte ptr [MAIN_NUMER+esp] ; load the numerator
fabs
rem1_large_loop:
fcom
fstsw ax
and eax, 00100h
jnz rem1_no_sub
fsub st, st(1)
rem1_no_sub:
fxch
fmul qword ptr half
fxch
sub ecx, 1 ; decrement the loop counter
jnz rem1_large_loop
mov ebx, [MAIN_NUMER+8+esp] ; sign and exponent of x (numerator)
fstp tbyte ptr[esp + MAIN_NUMER] ; save result
fstp st ; toss modified denom
fld tbyte ptr[esp + MAIN_DENOM_SAVE]
fld tbyte ptr[big_number] ; force C2 to be set
fprem1
fstp st
fld tbyte ptr[esp + MAIN_NUMER] ; restore saved result
fldcw [esp + MAIN_PREV_CW] ; restore caller's control word
and ebx, 08000h ; keep sx
jz rem1_done
fchs
jmp rem1_done
remainder1_hardware_ok:
fld tbyte ptr [MAIN_DENOM+esp] ; load the denominator
fld tbyte ptr [MAIN_NUMER+esp] ; load the numerator
fprem1 ; and finally do a remainder
; prem1_main_routine end
rem1_done:
test edx, 03h
jz rem1_exit
fnstsw [esp + MAIN_FPREM_SW] ; save Q0 Q1 and Q2
test edx, 01h
jz do_not_de_scale1
; De-scale the result. Go to pc=80 to prevent from fmul
; from user precision (fprem does not round the result).
fnstcw [esp + MAIN_PREV_CW] ; save callers control word
mov eax, [esp + MAIN_PREV_CW]
or eax, 0300h ; pc = 80
mov [esp + MAIN_PATCH_CW], eax
fldcw [esp + MAIN_PATCH_CW]
fmul qword ptr one_shr_64
fldcw [esp + MAIN_PREV_CW] ; restore callers CW
do_not_de_scale1:
mov eax, [esp + MAIN_FPREM_SW]
fxch
fstp st
fld tbyte ptr[esp + MAIN_DENOM_SAVE]
fxch
and eax, 04300h ; restore saved Q0, Q1, Q2
sub esp, ENV_SIZE
fnstenv [esp]
and [esp].STATUS_WORD, 0bcffh
or [esp].STATUS_WORD, eax
fldenv [esp]
add esp, ENV_SIZE
rem1_exit:
pop ecx
pop ebx
pop eax
CHECKSW ; debug only: save status
ret
fprem1_common ENDP
comment ~***************************************************************
;
; float frem1_chk (float numer, float denom)
;
public frem1_chk
frem1_chk PROC NEAR
push edx
sub esp, STACK_SIZE
fld dword ptr [STACK_SIZE+8+esp]
fstp tbyte ptr [NUMER+esp]
fld dword ptr [STACK_SIZE+12+esp]
fstp tbyte ptr [DENOM+esp]
mov edx, 0 ; dx = 1 if denormal extended divisor
call fprem1_common
fxch
fstp st
add esp, STACK_SIZE
pop edx
ret
frem1_chk ENDP
; end frem1_chk
;
; double drem1_chk (double numer, double denom)
;
public drem1_chk
drem1_chk PROC NEAR
push edx
sub esp, STACK_SIZE
fld qword ptr [STACK_SIZE+8+esp]
fstp tbyte ptr [NUMER+esp]
fld qword ptr [STACK_SIZE+16+esp]
fstp tbyte ptr [DENOM+esp]
mov edx, 0 ; dx = 1 if denormal extended divisor
call fprem1_common
fxch
fstp st
add esp, STACK_SIZE
pop edx
ret
drem1_chk ENDP
; end drem1_chk
;
; long double lrem1_chk(long double number,long double denom)
;
public lrem1_chk
lrem1_chk PROC NEAR
fld tbyte ptr [20+esp]
fld tbyte ptr [4+esp]
call fprem1_chk
fxch
fstp st
ret
lrem1_chk ENDP
********************************************************************~
;
; FPREM1: ST = remainder(ST, ST(1)) - IEEE version of rounding
;
; Compiler version of the FPREM must preserve the arguments in the floating
; point stack.
public __fprem1_chk
defpe __fprem1_chk
push edx
sub esp, STACK_SIZE
fstp tbyte ptr [NUMER+esp]
fstp tbyte ptr [DENOM+esp]
mov edx, 0
; prem1_main_routine begin
mov eax,[DENOM+6+esp] ; exponent and high 16 bits of mantissa
test eax,07fff0000h ; check for denormal
jz denormal1
call fprem1_common
add esp, STACK_SIZE
pop edx
ret
denormal1:
fld tbyte ptr [DENOM+esp] ; load the denominator
fld tbyte ptr [NUMER+esp] ; load the numerator
mov eax, [DENOM+esp] ; test for whole mantissa == 0
or eax, [DENOM+4+esp] ; test for whole mantissa == 0
jz remainder1_hardware_ok_l ; denominator is zero
fxch
fstp tbyte ptr[esp + DENOM_SAVE] ; save org denominator
fld tbyte ptr[esp + DENOM]
fxch
or edx, 02h
;
; For this we need pc=80. Also, mask exceptions so we don't take any
; denormal operand exceptions. It is guaranteed that the descaling
; later on will take underflow, which is what the hardware would have done
; on a normal fprem.
;
fnstcw [PREV_CW+esp] ; save caller's control word
mov eax, [PREV_CW+esp]
or eax, 0033fh ; mask exceptions, pc=80
mov [PATCH_CW+esp], eax
fldcw [PATCH_CW+esp] ; mask exceptions & pc=80
; The denominator is a denormal. For most numerators, scale both numerator
; and denominator to get rid of denormals. Then execute the common code
; with the flag set to indicate that the result must be de-scaled.
; For large numerators this won't work because the scaling would cause
; overflow. In this case we know the numerator is large, the denominator
; is small (denormal), so the exponent difference is also large. This means
; the rem1_large code will be used and this code depends on the difference
; in exponents modulo 64. Adding 64 to the denominators exponent
; doesn't change the modulo 64 difference. So we can scale the denominator
; by 64, making it not denormal, and this won't effect the result.
;
; To start with, figure out if numerator is large
mov eax, [esp + NUMER + 8] ; load numerator exponent
and eax, 7fffh ; isolate numerator exponent
cmp eax, 7fbeh ; compare Nexp to Maxexp-64
ja big_numer_rem1_de ; jif big numerator
; So the numerator is not large scale both numerator and denominator
or edx, 1 ; edx = 1, if denormal extended divisor
fmul qword ptr one_shl_64 ; make numerator not denormal
fstp tbyte ptr[esp + NUMER]
fmul qword ptr one_shl_64 ; make denominator not denormal
fstp tbyte ptr[esp + DENOM]
jmp scaling_done1
; The numerator is large. Scale only the denominator, which will not
; change the result which we know will be partial. Set the scale flag
; to false.
big_numer_rem1_de:
; We must do this with pc=80 to avoid rounding to single/double.
; In this case we do not mask exceptions so that we will take
; denormal operand, as would the hardware.
fnstcw [PREV_CW+esp] ; save caller's control word
mov eax, [PREV_CW+esp]
or eax, 00300h ; pc=80
mov [PATCH_CW+esp], eax
fldcw [PATCH_CW+esp] ; pc=80
fstp st ; Toss numerator
fmul qword ptr one_shl_64 ; make denominator not denormal
fstp tbyte ptr[esp + DENOM]
; Restore the control word which was fiddled to scale at 80-bit precision.
; Then call the common code.
scaling_done1:
fldcw [esp + PREV_CW] ; restore callers control word
call fprem1_common
add esp, STACK_SIZE
pop edx
ret
remainder1_hardware_ok_l:
fprem ; and finally do a remainder
CHECKSW
add esp, STACK_SIZE
pop edx
ret
__fprem1_chk ENDP
; end fprem1_chk
ifdef DEBUG
public fpinit
fpinit PROC NEAR
fninit
ret
fpinit ENDP
endif
_TEXT ENDS
END