pa-risc2w.s

mediastreamer2是开源的网络传输媒体流的库

;
; PA-RISC 64-bit implementation of bn_asm code
;
; This code is approximately 2x faster than the C version
; for RSA/DSA.
;
; See http://devresource.hp.com/  for more details on the PA-RISC
; architecture.  Also see the book "PA-RISC 2.0 Architecture"
; by Gerry Kane for information on the instruction set architecture.
;
; Code written by Chris Ruemmler (with some help from the HP C
; compiler).
;
; The code compiles with HP's assembler
;

	.level	2.0W
	.space	$TEXT$
	.subspa	$CODE$,QUAD=0,ALIGN=8,ACCESS=0x2c,CODE_ONLY

;
; Global Register definitions used for the routines.
;
; Some information about HP's runtime architecture for 64-bits.
;
; "Caller save" means the calling function must save the register
; if it wants the register to be preserved.
; "Callee save" means if a function uses the register, it must save
; the value before using it.
;
; For the floating point registers 
;
;    "caller save" registers: fr4-fr11, fr22-fr31
;    "callee save" registers: fr12-fr21
;    "special" registers: fr0-fr3 (status and exception registers)
;
; For the integer registers
;     value zero             :  r0
;     "caller save" registers: r1,r19-r26
;     "callee save" registers: r3-r18
;     return register        :  r2  (rp)
;     return values          ; r28  (ret0,ret1)
;     Stack pointer          ; r30  (sp) 
;     global data pointer    ; r27  (dp)
;     argument pointer       ; r29  (ap)
;     millicode return ptr   ; r31  (also a caller save register)


;
; Arguments to the routines
;
r_ptr       .reg %r26
a_ptr       .reg %r25
b_ptr       .reg %r24
num         .reg %r24
w           .reg %r23
n           .reg %r23


;
; Globals used in some routines
;

top_overflow .reg %r29
high_mask    .reg %r22    ; value 0xffffffff80000000L


;------------------------------------------------------------------------------
;
; bn_mul_add_words
;
;BN_ULONG bn_mul_add_words(BN_ULONG *r_ptr, BN_ULONG *a_ptr, 
;								int num, BN_ULONG w)
;
; arg0 = r_ptr
; arg1 = a_ptr
; arg2 = num
; arg3 = w
;
; Local register definitions
;

fm1          .reg %fr22
fm           .reg %fr23
ht_temp      .reg %fr24
ht_temp_1    .reg %fr25
lt_temp      .reg %fr26
lt_temp_1    .reg %fr27
fm1_1        .reg %fr28
fm_1         .reg %fr29

fw_h         .reg %fr7L
fw_l         .reg %fr7R
fw           .reg %fr7

fht_0        .reg %fr8L
flt_0        .reg %fr8R
t_float_0    .reg %fr8

fht_1        .reg %fr9L
flt_1        .reg %fr9R
t_float_1    .reg %fr9

tmp_0        .reg %r31
tmp_1        .reg %r21
m_0          .reg %r20 
m_1          .reg %r19 
ht_0         .reg %r1  
ht_1         .reg %r3
lt_0         .reg %r4
lt_1         .reg %r5
m1_0         .reg %r6 
m1_1         .reg %r7 
rp_val       .reg %r8
rp_val_1     .reg %r9

bn_mul_add_words
	.export	bn_mul_add_words,entry,NO_RELOCATION,LONG_RETURN
	.proc
	.callinfo frame=128
    .entry
	.align 64

    STD     %r3,0(%sp)          ; save r3  
    STD     %r4,8(%sp)          ; save r4  
	NOP                         ; Needed to make the loop 16-byte aligned
	NOP                         ; Needed to make the loop 16-byte aligned

    STD     %r5,16(%sp)         ; save r5  
    STD     %r6,24(%sp)         ; save r6  
    STD     %r7,32(%sp)         ; save r7  
    STD     %r8,40(%sp)         ; save r8  

    STD     %r9,48(%sp)         ; save r9  
    COPY    %r0,%ret0           ; return 0 by default
    DEPDI,Z 1,31,1,top_overflow ; top_overflow = 1 << 32    
	STD     w,56(%sp)           ; store w on stack

    CMPIB,>= 0,num,bn_mul_add_words_exit  ; if (num <= 0) then exit
	LDO     128(%sp),%sp       ; bump stack

	;
	; The loop is unrolled twice, so if there is only 1 number
    ; then go straight to the cleanup code.
	;
	CMPIB,= 1,num,bn_mul_add_words_single_top
	FLDD    -72(%sp),fw     ; load up w into fp register fw (fw_h/fw_l)

	;
	; This loop is unrolled 2 times (64-byte aligned as well)
	;
	; PA-RISC 2.0 chips have two fully pipelined multipliers, thus
    ; two 32-bit mutiplies can be issued per cycle.
    ; 
bn_mul_add_words_unroll2

    FLDD    0(a_ptr),t_float_0       ; load up 64-bit value (fr8L) ht(L)/lt(R)
    FLDD    8(a_ptr),t_float_1       ; load up 64-bit value (fr8L) ht(L)/lt(R)
    LDD     0(r_ptr),rp_val          ; rp[0]
    LDD     8(r_ptr),rp_val_1        ; rp[1]

    XMPYU   fht_0,fw_l,fm1           ; m1[0] = fht_0*fw_l
    XMPYU   fht_1,fw_l,fm1_1         ; m1[1] = fht_1*fw_l
    FSTD    fm1,-16(%sp)             ; -16(sp) = m1[0]
    FSTD    fm1_1,-48(%sp)           ; -48(sp) = m1[1]

    XMPYU   flt_0,fw_h,fm            ; m[0] = flt_0*fw_h
    XMPYU   flt_1,fw_h,fm_1          ; m[1] = flt_1*fw_h
    FSTD    fm,-8(%sp)               ; -8(sp) = m[0]
    FSTD    fm_1,-40(%sp)            ; -40(sp) = m[1]

    XMPYU   fht_0,fw_h,ht_temp       ; ht_temp   = fht_0*fw_h
    XMPYU   fht_1,fw_h,ht_temp_1     ; ht_temp_1 = fht_1*fw_h
    FSTD    ht_temp,-24(%sp)         ; -24(sp)   = ht_temp
    FSTD    ht_temp_1,-56(%sp)       ; -56(sp)   = ht_temp_1

    XMPYU   flt_0,fw_l,lt_temp       ; lt_temp = lt*fw_l
    XMPYU   flt_1,fw_l,lt_temp_1     ; lt_temp = lt*fw_l
    FSTD    lt_temp,-32(%sp)         ; -32(sp) = lt_temp 
    FSTD    lt_temp_1,-64(%sp)       ; -64(sp) = lt_temp_1 

    LDD     -8(%sp),m_0              ; m[0] 
    LDD     -40(%sp),m_1             ; m[1]
    LDD     -16(%sp),m1_0            ; m1[0]
    LDD     -48(%sp),m1_1            ; m1[1]

    LDD     -24(%sp),ht_0            ; ht[0]
    LDD     -56(%sp),ht_1            ; ht[1]
    ADD,L   m1_0,m_0,tmp_0           ; tmp_0 = m[0] + m1[0]; 
    ADD,L   m1_1,m_1,tmp_1           ; tmp_1 = m[1] + m1[1]; 

    LDD     -32(%sp),lt_0            
    LDD     -64(%sp),lt_1            
    CMPCLR,*>>= tmp_0,m1_0, %r0      ; if (m[0] < m1[0])
    ADD,L   ht_0,top_overflow,ht_0   ; ht[0] += (1<<32)

    CMPCLR,*>>= tmp_1,m1_1,%r0       ; if (m[1] < m1[1])
    ADD,L   ht_1,top_overflow,ht_1   ; ht[1] += (1<<32)
    EXTRD,U tmp_0,31,32,m_0          ; m[0]>>32  
    DEPD,Z  tmp_0,31,32,m1_0         ; m1[0] = m[0]<<32 

    EXTRD,U tmp_1,31,32,m_1          ; m[1]>>32  
    DEPD,Z  tmp_1,31,32,m1_1         ; m1[1] = m[1]<<32 
    ADD,L   ht_0,m_0,ht_0            ; ht[0]+= (m[0]>>32)
    ADD,L   ht_1,m_1,ht_1            ; ht[1]+= (m[1]>>32)

    ADD     lt_0,m1_0,lt_0           ; lt[0] = lt[0]+m1[0];
	ADD,DC  ht_0,%r0,ht_0            ; ht[0]++
    ADD     lt_1,m1_1,lt_1           ; lt[1] = lt[1]+m1[1];
    ADD,DC  ht_1,%r0,ht_1            ; ht[1]++

    ADD    %ret0,lt_0,lt_0           ; lt[0] = lt[0] + c;
	ADD,DC  ht_0,%r0,ht_0            ; ht[0]++
    ADD     lt_0,rp_val,lt_0         ; lt[0] = lt[0]+rp[0]
    ADD,DC  ht_0,%r0,ht_0            ; ht[0]++

	LDO    -2(num),num               ; num = num - 2;
    ADD     ht_0,lt_1,lt_1           ; lt[1] = lt[1] + ht_0 (c);
    ADD,DC  ht_1,%r0,ht_1            ; ht[1]++
    STD     lt_0,0(r_ptr)            ; rp[0] = lt[0]

    ADD     lt_1,rp_val_1,lt_1       ; lt[1] = lt[1]+rp[1]
    ADD,DC  ht_1,%r0,%ret0           ; ht[1]++
    LDO     16(a_ptr),a_ptr          ; a_ptr += 2

    STD     lt_1,8(r_ptr)            ; rp[1] = lt[1]
	CMPIB,<= 2,num,bn_mul_add_words_unroll2 ; go again if more to do
    LDO     16(r_ptr),r_ptr          ; r_ptr += 2

    CMPIB,=,N 0,num,bn_mul_add_words_exit ; are we done, or cleanup last one

	;
	; Top of loop aligned on 64-byte boundary
	;
bn_mul_add_words_single_top
    FLDD    0(a_ptr),t_float_0        ; load up 64-bit value (fr8L) ht(L)/lt(R)
    LDD     0(r_ptr),rp_val           ; rp[0]
    LDO     8(a_ptr),a_ptr            ; a_ptr++
    XMPYU   fht_0,fw_l,fm1            ; m1 = ht*fw_l
    FSTD    fm1,-16(%sp)              ; -16(sp) = m1
    XMPYU   flt_0,fw_h,fm             ; m = lt*fw_h
    FSTD    fm,-8(%sp)                ; -8(sp) = m
    XMPYU   fht_0,fw_h,ht_temp        ; ht_temp = ht*fw_h
    FSTD    ht_temp,-24(%sp)          ; -24(sp) = ht
    XMPYU   flt_0,fw_l,lt_temp        ; lt_temp = lt*fw_l
    FSTD    lt_temp,-32(%sp)          ; -32(sp) = lt 

    LDD     -8(%sp),m_0               
    LDD    -16(%sp),m1_0              ; m1 = temp1 
    ADD,L   m_0,m1_0,tmp_0            ; tmp_0 = m + m1; 
    LDD     -24(%sp),ht_0             
    LDD     -32(%sp),lt_0             

    CMPCLR,*>>= tmp_0,m1_0,%r0        ; if (m < m1)
    ADD,L   ht_0,top_overflow,ht_0    ; ht += (1<<32)

    EXTRD,U tmp_0,31,32,m_0           ; m>>32  
    DEPD,Z  tmp_0,31,32,m1_0          ; m1 = m<<32 

    ADD,L   ht_0,m_0,ht_0             ; ht+= (m>>32)
    ADD     lt_0,m1_0,tmp_0           ; tmp_0 = lt+m1;
    ADD,DC  ht_0,%r0,ht_0             ; ht++
    ADD     %ret0,tmp_0,lt_0          ; lt = lt + c;
    ADD,DC  ht_0,%r0,ht_0             ; ht++
    ADD     lt_0,rp_val,lt_0          ; lt = lt+rp[0]
    ADD,DC  ht_0,%r0,%ret0            ; ht++
    STD     lt_0,0(r_ptr)             ; rp[0] = lt

bn_mul_add_words_exit
    .EXIT
    LDD     -80(%sp),%r9              ; restore r9  
    LDD     -88(%sp),%r8              ; restore r8  
    LDD     -96(%sp),%r7              ; restore r7  
    LDD     -104(%sp),%r6             ; restore r6  
    LDD     -112(%sp),%r5             ; restore r5  
    LDD     -120(%sp),%r4             ; restore r4  
    BVE     (%rp)
    LDD,MB  -128(%sp),%r3             ; restore r3
	.PROCEND	;in=23,24,25,26,29;out=28;

;----------------------------------------------------------------------------
;
;BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
;
; arg0 = rp
; arg1 = ap
; arg2 = num
; arg3 = w

bn_mul_words
	.proc
	.callinfo frame=128
    .entry
	.EXPORT	bn_mul_words,ENTRY,PRIV_LEV=3,NO_RELOCATION,LONG_RETURN
	.align 64

    STD     %r3,0(%sp)          ; save r3  
    STD     %r4,8(%sp)          ; save r4  
    STD     %r5,16(%sp)         ; save r5  
    STD     %r6,24(%sp)         ; save r6  

    STD     %r7,32(%sp)         ; save r7  
    COPY    %r0,%ret0           ; return 0 by default
    DEPDI,Z 1,31,1,top_overflow ; top_overflow = 1 << 32    
	STD     w,56(%sp)           ; w on stack

    CMPIB,>= 0,num,bn_mul_words_exit
	LDO     128(%sp),%sp       ; bump stack

	;
	; See if only 1 word to do, thus just do cleanup
	;
	CMPIB,= 1,num,bn_mul_words_single_top
	FLDD    -72(%sp),fw     ; load up w into fp register fw (fw_h/fw_l)

	;
	; This loop is unrolled 2 times (64-byte aligned as well)
	;
	; PA-RISC 2.0 chips have two fully pipelined multipliers, thus
    ; two 32-bit mutiplies can be issued per cycle.
    ; 
bn_mul_words_unroll2

    FLDD    0(a_ptr),t_float_0        ; load up 64-bit value (fr8L) ht(L)/lt(R)
    FLDD    8(a_ptr),t_float_1        ; load up 64-bit value (fr8L) ht(L)/lt(R)
    XMPYU   fht_0,fw_l,fm1            ; m1[0] = fht_0*fw_l
    XMPYU   fht_1,fw_l,fm1_1          ; m1[1] = ht*fw_l

    FSTD    fm1,-16(%sp)              ; -16(sp) = m1
    FSTD    fm1_1,-48(%sp)            ; -48(sp) = m1
    XMPYU   flt_0,fw_h,fm             ; m = lt*fw_h
    XMPYU   flt_1,fw_h,fm_1           ; m = lt*fw_h

    FSTD    fm,-8(%sp)                ; -8(sp) = m
    FSTD    fm_1,-40(%sp)             ; -40(sp) = m
    XMPYU   fht_0,fw_h,ht_temp        ; ht_temp = fht_0*fw_h
    XMPYU   fht_1,fw_h,ht_temp_1      ; ht_temp = ht*fw_h

    FSTD    ht_temp,-24(%sp)          ; -24(sp) = ht
    FSTD    ht_temp_1,-56(%sp)        ; -56(sp) = ht
    XMPYU   flt_0,fw_l,lt_temp        ; lt_temp = lt*fw_l
    XMPYU   flt_1,fw_l,lt_temp_1      ; lt_temp = lt*fw_l

    FSTD    lt_temp,-32(%sp)          ; -32(sp) = lt 
    FSTD    lt_temp_1,-64(%sp)        ; -64(sp) = lt 
    LDD     -8(%sp),m_0               
    LDD     -40(%sp),m_1              

    LDD    -16(%sp),m1_0              
    LDD    -48(%sp),m1_1              
    LDD     -24(%sp),ht_0             
    LDD     -56(%sp),ht_1             

    ADD,L   m1_0,m_0,tmp_0            ; tmp_0 = m + m1; 
    ADD,L   m1_1,m_1,tmp_1            ; tmp_1 = m + m1; 
    LDD     -32(%sp),lt_0             
    LDD     -64(%sp),lt_1             

    CMPCLR,*>>= tmp_0,m1_0, %r0       ; if (m < m1)
    ADD,L   ht_0,top_overflow,ht_0    ; ht += (1<<32)
    CMPCLR,*>>= tmp_1,m1_1,%r0        ; if (m < m1)
    ADD,L   ht_1,top_overflow,ht_1    ; ht += (1<<32)

    EXTRD,U tmp_0,31,32,m_0           ; m>>32  
    DEPD,Z  tmp_0,31,32,m1_0          ; m1 = m<<32 
    EXTRD,U tmp_1,31,32,m_1           ; m>>32  
    DEPD,Z  tmp_1,31,32,m1_1          ; m1 = m<<32 

    ADD,L   ht_0,m_0,ht_0             ; ht+= (m>>32)
    ADD,L   ht_1,m_1,ht_1             ; ht+= (m>>32)
    ADD     lt_0,m1_0,lt_0            ; lt = lt+m1;
	ADD,DC  ht_0,%r0,ht_0             ; ht++

    ADD     lt_1,m1_1,lt_1            ; lt = lt+m1;
    ADD,DC  ht_1,%r0,ht_1             ; ht++
    ADD    %ret0,lt_0,lt_0            ; lt = lt + c (ret0);
	ADD,DC  ht_0,%r0,ht_0             ; ht++

    ADD     ht_0,lt_1,lt_1            ; lt = lt + c (ht_0)
    ADD,DC  ht_1,%r0,ht_1             ; ht++
    STD     lt_0,0(r_ptr)             ; rp[0] = lt
    STD     lt_1,8(r_ptr)             ; rp[1] = lt

	COPY    ht_1,%ret0                ; carry = ht
	LDO    -2(num),num                ; num = num - 2;
    LDO     16(a_ptr),a_ptr           ; ap += 2
	CMPIB,<= 2,num,bn_mul_words_unroll2
    LDO     16(r_ptr),r_ptr           ; rp++

    CMPIB,=,N 0,num,bn_mul_words_exit ; are we done?

	;
	; Top of loop aligned on 64-byte boundary
	;
bn_mul_words_single_top
    FLDD    0(a_ptr),t_float_0        ; load up 64-bit value (fr8L) ht(L)/lt(R)

    XMPYU   fht_0,fw_l,fm1            ; m1 = ht*fw_l
    FSTD    fm1,-16(%sp)              ; -16(sp) = m1
    XMPYU   flt_0,fw_h,fm             ; m = lt*fw_h
    FSTD    fm,-8(%sp)                ; -8(sp) = m
    XMPYU   fht_0,fw_h,ht_temp        ; ht_temp = ht*fw_h