1 files changed, 103 insertions, 0 deletions
diff --git a/modules/asm_amd64/mp_monty.S b/modules/asm_amd64/mp_monty.S
new file mode 100644
index 000000000..9c94d2736
--- /dev/null
+++ b/modules/asm_amd64/mp_monty.S
@@ -0,0 +1,103 @@
+/*************************************************
+* Montgomery Reduction Source File               *
+* (C) 2008 Jack Lloyd                            *
+*************************************************/
+
+#include <botan/asm_macr.h>
+
+START_LISTING(mp_monty.S)
+
+START_FUNCTION(bigint_monty_redc)
+
+#define Z_ARR    ARG_1 // rdi
+#define Z_SIZE   ARG_2_32 // esi
+#define X_ARR    ARG_3 // rdx
+#define X_SIZE   ARG_4_32 // ecx
+#define U        ARG_5 // r8
+
+/*
+     We need all arguments for a while (we can reuse U eventually)
+   So only temp registers are
+     TEMP_1 %r10
+     TEMP_2 %r11
+     TEMP_3 = ARG_6 = %r9
+   void return, so also
+     R0 %rax (aka TEMP_9)
+   is free
+
+   Can push:
+     %rbx (base pointer, callee saved)
+     %rpb (frame pointer, callee saved)
+     %r12-%r15 (callee saved)
+
+  Can push base/frame pointers since this is a leaf function
+  and does not reference any data.
+*/
+
+   push %r12
+   push %r13
+   push %r14
+   push %r15
+
+#define LOOP_CTR R0
+
+#define CARRY    TEMP_1
+#define Z_WORD   TEMP_2
+#define MUL_LO   %rax
+#define MUL_HI   %rdx
+
+   /*
+   ZEROIZE(CARRY)
+
+   ASSIGN(LOOP_CTR, X_SIZE)
+
+   JUMP_IF_ZERO(LOOP_CTR, .L_MULADD_DONE)
+   JUMP_IF_LT(LOOP_CTR, 8, .LOOP_MULADD1)
+
+#define MULADD_OP(N)                  \
+   ASSIGN(MUL_LO, ARRAY8(X_ARR, N)) ; \
+   ASSIGN(Z_WORD, ARRAY8(Z_ARR, N)) ; \
+   MUL(Y)                           ; \
+   ADD(Z_WORD, CARRY)               ; \
+   ASSIGN(CARRY, MUL_HI)            ; \
+   ADD_LAST_CARRY(CARRY)            ; \
+   ADD(Z_WORD, MUL_LO)              ; \
+   ADD_LAST_CARRY(CARRY)            ; \
+   ASSIGN(ARRAY8(Z_ARR, N), Z_WORD)
+
+ALIGN
+.LOOP_MULADD8:
+   MULADD_OP(0)
+   MULADD_OP(1)
+   MULADD_OP(2)
+   MULADD_OP(3)
+   MULADD_OP(4)
+   MULADD_OP(5)
+   MULADD_OP(6)
+   MULADD_OP(7)
+
+   SUB_IMM(LOOP_CTR, 8)
+   ADD_IMM(Z_ARR, 64)
+   ADD_IMM(X_ARR, 64)
+   cmp IMM(8), LOOP_CTR
+   jge .LOOP_MULADD8
+
+   JUMP_IF_ZERO(LOOP_CTR, .L_MULADD_DONE)
+
+ALIGN
+.LOOP_MULADD1:
+   MULADD_OP(0)
+
+   SUB_IMM(LOOP_CTR, 1)
+   ADD_IMM(Z_ARR, 8)
+   ADD_IMM(X_ARR, 8)
+
+   cmp IMM(0), LOOP_CTR
+   jne .LOOP_MULADD1
+*/
+
+   pop %r15
+   pop %r14
+   pop %r13
+   pop %r12
+END_FUNCTION(bigint_monty_redc)