;
; ARM Macros file
;
; Triple register is R4|R3|R2
; MUL_START. Initialise registers. Make R5 and R6 point to multipliers a 
; and b. R7 points at result c. 
; Initialise Triple register to 0
; See makemcs.txt for more information about this file
; 
;

MACRO PMUL_START
  __asm {
  MOV r5,a
  MOV r6,b
  MOV r7,c
  MOV r4,sn
  MOV r3,#0
  MOV r2,#0
ENDM

MACRO PMUL
  LDR r0,[r5,#(4*%d)]
  UMULL r8,r9,r0,r4
  ADDS r8,r8,r3
  ADC r3,r9,r2
  STR r2,[r6,#(4*%d)]
  STR r8,[r7,#(4*%d)]
ENDM

MACRO PMUL_END
  UMULL r8,r9,r3,r4
  STR r8,[r6]
  STR r9,[r6,#4]
  }
ENDM

MACRO H2_MUL_START
  __asm {
  MOV r5,a
  MOV r6,b
  MOV r7,c
  MOV r4,#0
  MOV r3,#0
  MOV r2,#0
  MOV r10,#0
  MOV r11,#0
ENDM

MACRO H2_STEP
    LDR r8,[r5,#(4*%d)]
    LDR r9,[r5,#(4*%d+4)]
    LDR r12,[r6,#(4*%d)]
    UMULL r0,r1,r8,r12
    ADDS r2,r2,r0
    ADCS r3,r3,r1
    ADC r11,r11,#0
    UMULL r0,r1,r9,r12
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    LDR r12,[r6,#(4*%d+4)]
    UMULL r0,r1,r8,r12
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    UMULL r0,r1,r9,r12
    ADDS r4,r4,r0
    ADCS r10,r10,r1
    ADDCS r11,r11,0x10000
ENDM

MACRO H2_LAST
    LDR r8,[r5,#(4*%d)]
    LDR r9,[r5,#(4*%d+4)]
    LDR r12,[r6,#(4*%d)]
    UMULL r0,r1,r8,r12
    ADDS r2,r2,r0
    ADCS r3,r3,r1
    UMULL r0,r1,r9,r12
    ADD r3,r3,r0
    LDR r12,[r6,#(4*%d+4)]
    UMULL r0,r1,r8,r12
    ADD r3,r3,r0
ENDM

MACRO H2_MFIN
     STR r2,[r7,#(4*%d)]
     STR r3,[r7,#(4*%d)]
     AND r0,r11,0xFF
     ADDS r2,r4,r0
     AND r0,r11,0xFF00
     ADCS r3,r10,r0,LSR#8
     ADDCS r11,r11,0x10000
     MOV r4,r11,LSR#16
     MOV r10,#0
     MOV r11,#0
ENDM

MACRO H2_MUL_END
   STR r2,[r7,#(4*%d)]
   STR r3,[r7,#(4*%d)] 
 }
ENDM

MACRO MUL_START
  __asm {
  MOV r5,a
  MOV r6,b
  MOV r7,c
  MOV r4,#0
  MOV r3,#0
  MOV r2,#0
ENDM
;
; STEP macro. Calculates a double-register partial product
; and adds it to the triple register total
; Parameters 1 & 2: Indices i and j for partial product multipliers a[i] 
; and b[j]
; STEP 1 even
MACRO STEP
  LDR r0,[r5,#(4*%d)]
  LDR r1,[r6,#(4*%d)]
  UMULL r8,r9,r0,r1
  ADDS r2,r2,r8
  ADCS r3,r3,r9
  ADC  r4,r4,#0
ENDM
;
; MFIN macro. Finish column calculation. Store Sum for this column
; and get Carry for next
; Parameter 1: Index k for Column Sum c[k]
MACRO MFIN
  STR r2,[r7,#(4*%d)]
  MOV r2,r3
  MOV r3,r4
  MOV r4,#0
ENDM
;
; LAST
;
MACRO LAST
  LDR r0,[r5,#(4*%d)]
  LDR r1,[r6,#(4*%d)]
  MLA r2,r0,r1,r2
ENDM
;
; MUL_END
; Parameter 1: Index for final carry c[.]
MACRO MUL_END  
  STR r2,[r7,#(4*%d)]
  }
ENDM

MACRO H2_SQR_START
  __asm {
  MOV r5,a
  MOV r7,c
  MOV r4,#0
  MOV r3,#0
  MOV r2,#0
  MOV r10,#0
  MOV r11,#0
ENDM

MACRO H2_DSTEP
    LDR r8,[r5,#(4*%d)]
    LDR r9,[r5,#(4*%d+4)]
    LDR r12,[r5,#(4*%d)]
    UMULL r0,r1,r8,r12
    ADDS r2,r2,r0
    ADCS r3,r3,r1
    ADC r11,r11,#0
    ADDS r2,r2,r0
    ADCS r3,r3,r1
    ADC r11,r11,#0
    UMULL r0,r1,r9,r12
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    LDR r12,[r5,#(4*%d+4)]
    UMULL r0,r1,r8,r12
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    UMULL r0,r1,r9,r12
    ADDS r4,r4,r0
    ADCS r10,r10,r1
    ADDCS r11,r11,0x10000
    ADDS r4,r4,r0
    ADCS r10,r10,r1
    ADDCS r11,r11,0x10000
ENDM

MACRO H2_SELF
    LDR r8,[r5,#(4*%d)]
    LDR r9,[r5,#(4*%d+4)]
    UMULL r0,r1,r8,r8
    ADDS r2,r2,r0
    ADCS r3,r3,r1
    ADC r11,r11,#0
    UMULL r0,r1,r8,r9
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    ADDS r3,r3,r0
    ADCS r4,r4,r1
    ADDCS r11,r11,0x100
    UMULL r0,r1,r9,r9
    ADDS r4,r4,r0
    ADCS r10,r10,r1
    ADDCS r11,r11,0x10000
ENDM

MACRO H2_SFIN
     STR r2,[r7,#(4*%d)]
     STR r3,[r7,#(4*%d)]
     AND r0,r11,0xFF
     ADDS r2,r4,r0
     AND r0,r11,0xFF00
     ADCS r3,r10,r0,LSR#8
     ADDCS r11,r11,0x10000
     MOV r4,r11,LSR#16
     MOV r10,#0
     MOV r11,#0
ENDM

MACRO H2_SQR_END
   STR r2,[r7,#(4*%d)]
   STR r3,[r7,#(4*%d)] 
   }
ENDM

;
; SQR_START
; 
MACRO SQR_START
  __asm {
  MOV r5,a
  MOV r7,c
  MOV r4,#0
  MOV r3,#0
  MOV r2,#0
ENDM
;
; DSTEP macro. Calculates a double-register partial product
; and add it twice to a triple register total
; Parameters 1 & 2 : Indices of partial product multipliers
MACRO DSTEP
  LDR  r0,[r5,#(4*%d)]
  LDR  r1,[r5,#(4*%d)]
  UMULL r8,r9,r0,r1
  ADDS r2,r2,r8
  ADCS r3,r3,r9
  ADC  r4,r4,#0
  ADDS r2,r2,r8
  ADCS r3,r3,r9
  ADC  r4,r4,#0
ENDM

;                
; SELF macro. Calculate the double-register square and
; add it to a triple register total
; Parameter 1 : Index of diagonal element
MACRO SELF
  LDR r0,[r5,#(4*%d)]
  UMULL r8,r9,r0,r0
  ADDS r2,r2,r8
  ADCS r3,r3,r9
  ADC  r4,r4,#0
ENDM
;
; SFIN macro. Finish column calculation for squaring. Store Sum
; and get Carry for next column.
; Parameter 1: Index of Column Sum
MACRO SFIN
  STR r2,[r7,#(4*%d)]
  MOV r2,r3
  MOV r3,r4
  MOV r4,#0
ENDM
;
; SQR_END
; Parameter 1: Index for final carry
MACRO SQR_END
  STR r2,[r7,#(4*%d)]
  }
ENDM
;
; REDC_START macro
;
MACRO REDC_START
  __asm {
  MOV r5,a
  MOV r6,b
  MOV r7,ndash
  MOV r4,#0
  MOV r3,#0
  LDR r2,[r5]
ENDM
;
; RFINU macro
;
MACRO RFINU
  MUL r1,r7,r2  
  STR r1,[r5,#(4*%d)]
  LDR r0,[r6]
  UMULL r8,r9,r0,r1
  ADDS r0,r2,r8
  ADCS r2,r3,r9
  ADC  r3,r4,#0
  LDR  r0,[r5,#(4*(%d+1))]
  ADDS r2,r2,r0
  ADC  r3,r3,#0
  MOV  r4,#0
ENDM
;
; RFIND macro
;
MACRO RFIND
  STR  r2,[r5,#(4*%d)]
  LDR  r0,[r5,#(4*(%d+1))]
  ADDS r2,r3,r0
  ADC  r3,r4,#0
  MOV  r4,#0  
ENDM
;
; REDC_END
;
MACRO REDC_END
  STR  r2,[r5,#(4*%d)]
  STR  r3,[r5,#(4*(%d+1))]
  }
ENDM
;
; ADD_START macro - initialise for add/subtract, do first one
;
MACRO ADD_START
  __asm {
  MOV  r6,a
  MOV  r7,b
  MOV  r8,c
  LDR  r0,[r6]
  LDR  r1,[r7]
  ADDS r0,r0,r1
  STR  r0,[r8]
ENDM
;
; ADD macro. Add two numbers from memory and store result in memory.
; Don't forget carry bit
;
MACRO ADD
  LDR  r0,[r6,#(4*%d)]
  LDR  r1,[r7,#(4*%d)]
  ADCS r0,r0,r1
  STR  r0,[r8,#(4*%d)]
ENDM
;
; ADD_END macro. Catch carry
;
MACRO ADD_END
  MOV  r0,#0
  MOVCS r0,#1
  MOV carry,r0
  }
ENDM
;
; INC_START macro. Do first one
;
MACRO INC_START
  __asm {
  MOV  r6,a
  MOV  r7,b
  LDR  r0,[r6]
  LDR  r1,[r7]
  ADDS r0,r0,r1
  STR  r0,[r6]
ENDM
;
; INC macro. Add two numbers from memory and store result in memory.
; Don't forget carry bit
;
MACRO INC
  LDR  r0,[r6,#(4*%d)]
  LDR  r1,[r7,#(4*%d)]
  ADCS r0,r0,r1
  STR  r0,[r6,#(4*%d)]
ENDM
;
; INC_END macro. Catch carry
;
MACRO INC_END
  MOV   r0,#0
  MOVCS r0,#1
  MOV carry,r0
  }
ENDM
;
; SUB_START macro
;
MACRO SUB_START
__asm {
  MOV r6,a
  MOV r7,b
  MOV r8,c
  LDR  r0,[r6]
  LDR  r1,[r7]
  SUBS r0,r0,r1
  STR  r0,[r8]
ENDM
;
; SUB macro. Subtract two numbers in memory and store result in memory.
;
MACRO SUB
  LDR  r0,[r6,#(4*%d)]
  LDR  r1,[r7,#(4*%d)]
  SBCS r0,r0,r1
  STR  r0,[r8,#(4*%d)]
ENDM
;
; SUB_END macro. Catch carry
;
MACRO SUB_END
  MOV r0,#0
  MOVCC r0,#1
  MOV carry,r0
  }
ENDM
;
; DEC_START macro
;
MACRO DEC_START
__asm {
  MOV r6,a
  MOV r7,b
  LDR  r0,[r6]
  LDR  r1,[r7]
  SUBS r0,r0,r1
  STR  r0,[r6]
ENDM
;
; DEC macro. Subtract two numbers in memory and store result in memory.
;
MACRO DEC
  LDR  r0,[r6,#(4*%d)]
  LDR  r1,[r7,#(4*%d)]
  SBCS r0,r0,r1
  STR  r0,[r6,#(4*%d)]
ENDM
;
; DEC_END macro. Catch carry
;
MACRO DEC_END
  MOV r0,#0
  MOVCC r0,#1
  MOV carry,r0
  }
ENDM
;
; KADD_START macro. Zero Carry flag
;
MACRO KADD_START
  __asm {
  MOV r6,a
  MOV r7,b
  MOV r8,c
  MOV r9,n
  MOV r0,#0
  ADDS r0,r0,r0
  k%d:
ENDM
;
; KASL macro. Important that carry flag is undisturbed!
;
MACRO KASL
  SUB r9,r9,#1
  TEQ r9,#0
  BEQ k%d
  ADD r6,r6,#(4*%d)
  ADD r7,r7,#(4*%d)
  ADD r8,r8,#(4*%d)
  B k%d
  k%d:
ENDM
;
; KADD_END macro
;
MACRO KADD_END
  MOV r0,#0
  MOVCS r0,#1
  MOV carry,r0
  }
ENDM
;
; KINC_START macro. Set carry to Zero 
;
MACRO KINC_START
  __asm {
  MOV r6,a
  MOV r7,b
  MOV r9,n
  MOV r0,#0
  ADDS r0,r0,r0
  k%d:
ENDM
;
; KIDL macro. Important that carry flag is undisturbed!
;
MACRO KIDL
  SUB r9,r9,#1
  TEQ r9,#0
  BEQ k%d
  ADD r6,r6,#(4*%d)
  ADD r7,r7,#(4*%d)
  B k%d
  k%d:
ENDM
;
; KINC_END macro
;
MACRO KINC_END
  MOV r0,#0
  MOVCS r0,#1
  MOV carry,r0
  }
ENDM
;
; KDEC_START macro. Set carry
;
MACRO KDEC_START
  __asm {
  MOV r6,a
  MOV r7,b
  MOV r9,n
  SUBS r0,r0,r0
  k%d:
ENDM
;
; KDEC_END macro
;
MACRO KDEC_END
  MOV r0,#0
  MOVCC r0,#1
  MOV carry,r0
  }
ENDM