Initial x86-64 assembly code for the inner multiply-add loop.

3 files changed, 175 insertions, 0 deletions

diff --git a/modules/alg_amd64/asm_macr.h b/modules/alg_amd64/asm_macr.h
new file mode 100644
index 000000000..689bfa3cb
--- /dev/null
+++ b/modules/alg_amd64/asm_macr.h
@@ -0,0 +1,81 @@
+/*************************************************
+* Assembly Macros Header File                    *
+* (C) 1999-2006 The Botan Project                *
+*************************************************/
+
+#ifndef BOTAN_EXT_ASM_MACROS_H__
+#define BOTAN_EXT_ASM_MACROS_H__
+
+/*************************************************
+* General/Global Macros                          *
+*************************************************/
+#define ALIGN .p2align 4,,15
+
+#define START_LISTING(FILENAME) \
+   .file #FILENAME;             \
+   .text;                       \
+   ALIGN;
+
+/*************************************************
+* Function Definitions                           *
+*************************************************/
+#define START_FUNCTION(func_name) \
+   ALIGN;                         \
+   .global  func_name;            \
+   .type    func_name,@function;  \
+func_name:
+
+#define END_FUNCTION(func_name) \
+   ret
+
+/*************************************************
+* Loop Control                                   *
+*************************************************/
+#define START_LOOP(LABEL) \
+   ALIGN;                 \
+   LABEL##_LOOP:
+
+#define LOOP_UNTIL_EQ(REG, NUM, LABEL) \
+   cmp IMM(NUM), REG;                  \
+   jne LABEL##_LOOP
+
+#define LOOP_UNTIL_LT(REG, NUM, LABEL) \
+   cmp IMM(NUM), REG;                  \
+   jge LABEL##_LOOP
+
+/*************************************************
+ Conditional Jumps                              *
+*************************************************/
+#define JUMP_IF_ZERO(REG, LABEL) \
+   cmp IMM(0), REG;              \
+   jz LABEL
+
+#define JUMP_IF_LT(REG, NUM, LABEL) \
+   cmp IMM(NUM), REG;               \
+   jl LABEL
+
+/*************************************************
+* Memory Access Operations                       *
+*************************************************/
+#define ARRAY8(REG, NUM) 8*(NUM)(REG)
+
+#define ASSIGN(TO, FROM) mov FROM, TO
+
+/*************************************************
+* ALU Operations                                 *
+*************************************************/
+#define IMM(VAL) $VAL
+
+#define ADD(TO, FROM) addq FROM, TO
+#define ADD_IMM(TO, NUM) ADD(TO, IMM(NUM))
+#define ADD_W_CARRY(TO1, TO2, FROM) addq FROM, TO1; adcq IMM(0), TO2;
+#define SUB_IMM(TO, NUM) sub IMM(NUM), TO
+#define MUL(REG) mulq REG
+
+#define XOR(TO, FROM) xorq FROM, TO
+#define AND(TO, FROM) andq FROM, TO
+#define OR(TO, FROM) orq FROM, TO
+#define NOT(REG) notq REG
+#define ZEROIZE(REG) XOR(REG, REG)
+
+#endif
diff --git a/modules/alg_amd64/modinfo.txt b/modules/alg_amd64/modinfo.txt
new file mode 100644
index 000000000..5f1a98ea3
--- /dev/null
+++ b/modules/alg_amd64/modinfo.txt
@@ -0,0 +1,34 @@
+realname "Algorithm x86-64 Assembler"
+
+mp_bits 64
+
+<replace>
+#sha160.cpp
+</replace>
+
+<ignore>
+mp_muladd.cpp
+</ignore>
+
+<add>
+asm_macr.h
+mp_muladd.S
+#sha1core.S
+</add>
+
+<arch>
+amd64
+</arch>
+
+<cc>
+gcc
+</cc>
+
+# ELF systems
+<os>
+linux
+freebsd
+netbsd
+openbsd
+solaris
+</os>
diff --git a/modules/alg_amd64/mp_muladd.S b/modules/alg_amd64/mp_muladd.S
new file mode 100644
index 000000000..e34b855bb
--- /dev/null
+++ b/modules/alg_amd64/mp_muladd.S
@@ -0,0 +1,60 @@
+/*************************************************
+* Multiply/Add Algorithm Source File             *
+* (C) 1999-2006 The Botan Project                *
+*************************************************/
+
+#include <botan/asm_macr.h>
+
+START_LISTING(mp_muladd.S)
+
+START_FUNCTION(bigint_mul_add_words)
+#define LOOP_CTR %r11d
+#define X_ARR %rsi
+#define Z_ARR %rdi
+#define Y %rcx
+#define CARRY %r8
+
+   ZEROIZE(CARRY)
+
+   ASSIGN(LOOP_CTR, %edx)
+
+   JUMP_IF_ZERO(LOOP_CTR, .DONE)
+   JUMP_IF_LT(LOOP_CTR, 8, .MULADD1_LOOP)
+
+#define MULADD_OP(N)                \
+   ASSIGN(%rax, ARRAY8(X_ARR, N)) ; \
+   ASSIGN(%r9,  ARRAY8(Z_ARR, N)) ; \
+   MUL(Y)                         ; \
+   ADD_W_CARRY(%rax, %rdx, CARRY) ; \
+   ADD_W_CARRY(%rax, %rdx, %r9)   ; \
+   ASSIGN(CARRY, %rdx)            ; \
+   ASSIGN(ARRAY8(Z_ARR, N), %rax)
+
+START_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)
+LOOP_UNTIL_LT(LOOP_CTR, 8, .MULADD8)
+
+   JUMP_IF_ZERO(LOOP_CTR, .DONE)
+
+START_LOOP(.MULADD1)
+   MULADD_OP(0)
+
+   SUB_IMM(LOOP_CTR, 1)
+   ADD_IMM(Z_ARR, 8)
+   ADD_IMM(X_ARR, 8)
+LOOP_UNTIL_EQ(LOOP_CTR, 0, .MULADD1)
+
+.DONE:
+   ASSIGN(%rax, CARRY)
+END_FUNCTION(bigint_mul_add_words)