Initial checkin1.5.6

3 files changed, 273 insertions, 0 deletions

diff --git a/modules/mp_amd64/modinfo.txt b/modules/mp_amd64/modinfo.txt
new file mode 100644
index 000000000..12537da0e
--- /dev/null
+++ b/modules/mp_amd64/modinfo.txt
@@ -0,0 +1,15 @@
+realname "MPI Assembler Backend for x86-64/AMD64 Systems"
+
+mp_bits 64
+
+replace_file mp_asm.h
+replace_file mp_asmi.h
+
+<arch>
+amd64
+</arch>
+
+<cc>
+gcc
+icc
+</cc>
diff --git a/modules/mp_amd64/mp_asm.h b/modules/mp_amd64/mp_asm.h
new file mode 100644
index 000000000..17f33224c
--- /dev/null
+++ b/modules/mp_amd64/mp_asm.h
@@ -0,0 +1,84 @@
+/*************************************************
+* Lowest Level MPI Algorithms Header File        *
+* (C) 1999-2006 The Botan Project                *
+*************************************************/
+
+#ifndef BOTAN_MP_ASM_H__
+#define BOTAN_MP_ASM_H__
+
+#include <botan/mp_types.h>
+
+#if (BOTAN_MP_WORD_BITS != 64)
+   #error The mp_amd64 module requires that BOTAN_MP_WORD_BITS == 64
+#endif
+
+namespace Botan {
+
+extern "C" {
+
+/*************************************************
+* Word Multiply                                  *
+*************************************************/
+inline word word_madd2(word a, word b, word c, word* carry)
+   {
+   asm(
+      "mulq %1\n\t"        // a(in eax) * b(wherever) -> edx:eax
+      "addq %5,%0\n\t"     // add c to low word (eax)
+      "adcq $0,%2"         // add carry from previous to high word (edx)
+      : "=a"(a), "=rm"(b), "=&d"(*carry)
+      : "0"(a), "1"(b), "g"(c) : "cc");
+   return a;
+   }
+
+/*************************************************
+* Word Multiply/Add                              *
+*************************************************/
+inline word word_madd3(word a, word b, word c, word d, word* carry)
+   {
+   asm(
+      "mulq %1\n\t"        // a(in eax) * b(wherever) -> edx:eax
+      "addq %5,%0\n\t"     // add c to low word (eax)
+      "adcq $0,%2\n\t"     // add carry from previous to high word (edx)
+      "addq %6,%0\n\t"     // add d to low word (eax)
+      "adcq $0,%2"         // add carry from previous to high word (edx)
+      : "=a"(a), "=rm"(b), "=&d"(*carry)
+      : "0"(a), "1"(b), "g"(c), "g"(d) : "cc");
+   return a;
+   }
+
+/*************************************************
+* Multiply-Add Accumulator                       *
+*************************************************/
+inline void word3_muladd(word* w2, word* w1, word* w0, word a, word b)
+   {
+   asm("mulq %[b]\n\t"        // a(in eax) * b(wherever) -> edx:eax
+       "addq %3,%[w0]\n\t"     // add c to low word (eax)
+       "adcq %4,%[w1]\n\t"     // add carry from previous to high word (edx)
+       "adcq $0,%[w2]\n\t"     // add carry from previous to high word (edx)
+       : [w0]"=r"(*w0), [w1]"=r"(*w1), [w2]"=r"(*w2)
+       : "a"(a), [b]"d"(b), "0"(*w0), "1"(*w1), "2"(*w2)
+       : "cc");
+   }
+
+/*************************************************
+* Multiply-Add Accumulator                       *
+*************************************************/
+inline void word3_muladd_2(word* w2, word* w1, word* w0, word a, word b)
+   {
+   asm("mulq %[b]\n\t"        // a(in eax) * b(wherever) -> edx:eax
+       "addq %3,%[w0]\n\t"     // add c to low word (eax)
+       "adcq %4,%[w1]\n\t"     // add carry from previous to high word (edx)
+       "adcq $0,%[w2]\n\t"     // add carry from previous to high word (edx)
+       "addq %3,%[w0]\n\t"     // add c to low word (eax)
+       "adcq %4,%[w1]\n\t"     // add carry from previous to high word (edx)
+       "adcq $0,%[w2]\n\t"     // add carry from previous to high word (edx)
+       : [w0]"=r"(*w0), [w1]"=r"(*w1), [w2]"=r"(*w2)
+       : "a"(a), [b]"d"(b), "0"(*w0), "1"(*w1), "2"(*w2)
+       : "cc");
+      }
+
+}
+
+}
+
+#endif
diff --git a/modules/mp_amd64/mp_asmi.h b/modules/mp_amd64/mp_asmi.h
new file mode 100644
index 000000000..daa1243ca
--- /dev/null
+++ b/modules/mp_amd64/mp_asmi.h
@@ -0,0 +1,174 @@
+/*************************************************
+* Lowest Level MPI Algorithms Header File        *
+* (C) 1999-2006 The Botan Project                *
+*************************************************/
+
+#ifndef BOTAN_MP_ASM_INTERNAL_H__
+#define BOTAN_MP_ASM_INTERNAL_H__
+
+#include <botan/mp_types.h>
+
+namespace Botan {
+
+/*************************************************
+* Helper Macros for amd64 Assembly               *
+*************************************************/
+#define ASM(x) x "\n\t"
+
+#define ADDSUB2_OP(OPERATION, INDEX)                    \
+        ASM("movq 8*" INDEX "(%[y]), %[carry]")         \
+        ASM(OPERATION " %[carry], 8*" INDEX "(%[x])")   \
+
+#define ADDSUB3_OP(OPERATION, INDEX)                    \
+        ASM("movq 8*" INDEX "(%[x]), %[carry]")         \
+        ASM(OPERATION " 8*" INDEX "(%[y]), %[carry]")   \
+        ASM("movq %[carry], 8*" INDEX "(%[z])")         \
+
+#define LINMUL_OP(WRITE_TO, INDEX)                      \
+        ASM("movq 8*" INDEX "(%[x]),%%rax")             \
+        ASM("mulq %[y]")                                \
+        ASM("addq %[carry],%%rax")                      \
+        ASM("adcq $0,%%rdx")                            \
+        ASM("movq %%rdx,%[carry]")                      \
+        ASM("movq %%rax, 8*" INDEX "(%[" WRITE_TO "])")
+
+#define MULADD_OP(IGNORED, INDEX)                       \
+        ASM("movq 8*" INDEX "(%[x]),%%rax")             \
+        ASM("mulq %[y]")                                \
+        ASM("addq %[carry],%%rax")                      \
+        ASM("adcq $0,%%rdx")                            \
+        ASM("addq 8*" INDEX "(%[z]),%%rax")             \
+        ASM("adcq $0,%%rdx")                            \
+        ASM("movq %%rdx,%[carry]")                      \
+        ASM("movq %%rax, 8*" INDEX " (%[z])")
+
+#define DO_8_TIMES(MACRO, ARG) \
+        MACRO(ARG, "0") MACRO(ARG, "1") MACRO(ARG, "2") MACRO(ARG, "3") \
+        MACRO(ARG, "4") MACRO(ARG, "5") MACRO(ARG, "6") MACRO(ARG, "7")
+
+#define ADD_OR_SUBTRACT(CORE_CODE)     \
+        ASM("rorq %[carry]")           \
+        CORE_CODE                      \
+        ASM("sbbq %[carry],%[carry]")  \
+        ASM("negq %[carry]")
+
+/*************************************************
+* Word Addition                                  *
+*************************************************/
+inline word word_add(word x, word y, word* carry)
+   {
+   asm(
+      ADD_OR_SUBTRACT(ASM("adcq %[y],%[x]"))
+      : [x]"=r"(x), [carry]"=r"(*carry)
+      : "0"(x), [y]"rm"(y), "1"(*carry)
+      : "cc");
+   return x;
+   }
+
+/*************************************************
+* Eight Word Block Addition, Two Argument        *
+*************************************************/
+inline word word8_add2(word x[8], const word y[8], word carry)
+   {
+   asm(
+      ADD_OR_SUBTRACT(DO_8_TIMES(ADDSUB2_OP, "adcq"))
+      : [carry]"=r"(carry)
+      : [x]"r"(x), [y]"r"(y), "0"(carry)
+      : "cc", "memory");
+   return carry;
+   }
+
+/*************************************************
+* Eight Word Block Addition, Three Argument      *
+*************************************************/
+inline word word8_add3(word z[8], const word x[8], const word y[8], word carry)
+   {
+   asm(
+      ADD_OR_SUBTRACT(DO_8_TIMES(ADDSUB3_OP, "adcq"))
+      : [carry]"=r"(carry)
+      : [x]"r"(x), [y]"r"(y), [z]"r"(z), "0"(carry)
+      : "cc", "memory");
+   return carry;
+   }
+
+/*************************************************
+* Word Subtraction                               *
+*************************************************/
+inline word word_sub(word x, word y, word* carry)
+   {
+   asm(
+      ADD_OR_SUBTRACT(ASM("sbbq %[y],%[x]"))
+      : [x]"=r"(x), [carry]"=r"(*carry)
+      : "0"(x), [y]"rm"(y), "1"(*carry)
+      : "cc");
+   return x;
+   }
+
+/*************************************************
+* Eight Word Block Subtraction, Two Argument     *
+*************************************************/
+inline word word8_sub2(word x[8], const word y[8], word carry)
+   {
+   asm(
+      ADD_OR_SUBTRACT(DO_8_TIMES(ADDSUB2_OP, "sbbq"))
+      : [carry]"=r"(carry)
+      : [x]"r"(x), [y]"r"(y), "0"(carry)
+      : "cc", "memory");
+   return carry;
+   }
+
+/*************************************************
+* Eight Word Block Subtraction, Three Argument   *
+*************************************************/
+inline word word8_sub3(word z[8], const word x[8], const word y[8], word carry)
+   {
+   asm(
+      ADD_OR_SUBTRACT(DO_8_TIMES(ADDSUB3_OP, "sbbq"))
+      : [carry]"=r"(carry)
+      : [x]"r"(x), [y]"r"(y), [z]"r"(z), "0"(carry)
+      : "cc", "memory");
+   return carry;
+   }
+
+/*************************************************
+* Eight Word Block Linear Multiplication         *
+*************************************************/
+inline word word8_linmul2(word x[8], word y, word carry)
+   {
+   asm(
+      DO_8_TIMES(LINMUL_OP, "x")
+      : [carry]"=r"(carry)
+      : [x]"r"(x), [y]"rm"(y), "0"(carry)
+      : "cc", "%rax", "%rdx");
+   return carry;
+   }
+
+/*************************************************
+* Eight Word Block Linear Multiplication          *
+*************************************************/
+inline word word8_linmul3(word z[8], const word x[8], word y, word carry)
+   {
+   asm(
+      DO_8_TIMES(LINMUL_OP, "z")
+      : [carry]"=r"(carry)
+      : [z]"r"(z), [x]"r"(x), [y]"rm"(y), "0"(carry)
+      : "cc", "%rax", "%rdx");
+   return carry;
+   }
+
+/*************************************************
+* Eight Word Block Multiply/Add                  *
+*************************************************/
+inline word word8_madd3(word z[8], const word x[8], word y, word carry)
+   {
+   asm(
+      DO_8_TIMES(MULADD_OP, "")
+      : [carry]"=r"(carry)
+      : [z]"r"(z), [x]"r"(x), [y]"rm"(y), "0"(carry)
+      : "cc", "%rax", "%rdx");
+   return carry;
+   }
+
+}
+
+#endif