Enable AMD64 Twofish implementation on WIN64

* cipher/twofish-amd64.S: Enable when
HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS defined.
(ELF): New macro to mask lines with ELF specific commands.
* cipher/twofish.c (USE_AMD64_ASM): Enable when
HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS defined.
[HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS] (call_sysv_fn): New.
(twofish_amd64_encrypt_block, twofish_amd64_decrypt_block)
(twofish_amd64_ctr_enc, twofish_amd64_cbc_dec)
(twofish_amd64_cfb_dec): New wrapper functions for AMD64
assembly functions.
--

Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>

2 files changed, 100 insertions, 21 deletions

diff --git a/cipher/twofish-amd64.S b/cipher/twofish-amd64.S
index a2253076..ea88b94e 100644
--- a/cipher/twofish-amd64.S
+++ b/cipher/twofish-amd64.S
@@ -20,7 +20,14 @@
 
 #ifdef __x86_64
 #include <config.h>
-#if defined(HAVE_COMPATIBLE_GCC_AMD64_PLATFORM_AS) && defined(USE_TWOFISH)
+#if (defined(HAVE_COMPATIBLE_GCC_AMD64_PLATFORM_AS) || \
+    defined(HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS)) && defined(USE_TWOFISH)
+
+#ifdef HAVE_COMPATIBLE_GCC_AMD64_PLATFORM_AS
+# define ELF(...) __VA_ARGS__
+#else
+# define ELF(...) /*_*/
+#endif
 
 #ifdef __PIC__
 #  define RIP %rip
@@ -166,7 +173,7 @@
 
 .align 8
 .globl _gcry_twofish_amd64_encrypt_block
-.type   _gcry_twofish_amd64_encrypt_block,@function;
+ELF(.type   _gcry_twofish_amd64_encrypt_block,@function;)
 
 _gcry_twofish_amd64_encrypt_block:
 	/* input:
@@ -205,11 +212,11 @@ _gcry_twofish_amd64_encrypt_block:
 	addq $(3 * 8), %rsp;
 
 	ret;
-.size _gcry_twofish_amd64_encrypt_block,.-_gcry_twofish_amd64_encrypt_block;
+ELF(.size _gcry_twofish_amd64_encrypt_block,.-_gcry_twofish_amd64_encrypt_block;)
 
 .align 8
 .globl _gcry_twofish_amd64_decrypt_block
-.type   _gcry_twofish_amd64_decrypt_block,@function;
+ELF(.type   _gcry_twofish_amd64_decrypt_block,@function;)
 
 _gcry_twofish_amd64_decrypt_block:
 	/* input:
@@ -248,7 +255,7 @@ _gcry_twofish_amd64_decrypt_block:
 	addq $(3 * 8), %rsp;
 
 	ret;
-.size _gcry_twofish_amd64_encrypt_block,.-_gcry_twofish_amd64_encrypt_block;
+ELF(.size _gcry_twofish_amd64_encrypt_block,.-_gcry_twofish_amd64_encrypt_block;)
 
 #undef CTX
 
@@ -462,7 +469,7 @@ _gcry_twofish_amd64_decrypt_block:
 	outunpack3(RAB, 2);
 
 .align 8
-.type __twofish_enc_blk3,@function;
+ELF(.type __twofish_enc_blk3,@function;)
 
 __twofish_enc_blk3:
 	/* input:
@@ -485,10 +492,10 @@ __twofish_enc_blk3:
 	outunpack_enc3();
 
 	ret;
-.size __twofish_enc_blk3,.-__twofish_enc_blk3;
+ELF(.size __twofish_enc_blk3,.-__twofish_enc_blk3;)
 
 .align 8
-.type  __twofish_dec_blk3,@function;
+ELF(.type  __twofish_dec_blk3,@function;)
 
 __twofish_dec_blk3:
 	/* input:
@@ -511,11 +518,11 @@ __twofish_dec_blk3:
 	outunpack_dec3();
 
 	ret;
-.size __twofish_dec_blk3,.-__twofish_dec_blk3;
+ELF(.size __twofish_dec_blk3,.-__twofish_dec_blk3;)
 
 .align 8
 .globl _gcry_twofish_amd64_ctr_enc
-.type   _gcry_twofish_amd64_ctr_enc,@function;
+ELF(.type   _gcry_twofish_amd64_ctr_enc,@function;)
 _gcry_twofish_amd64_ctr_enc:
 	/* input:
 	 *	%rdi: ctx, CTX
@@ -593,11 +600,11 @@ _gcry_twofish_amd64_ctr_enc:
 	addq $(8 * 8), %rsp;
 
 	ret;
-.size _gcry_twofish_amd64_ctr_enc,.-_gcry_twofish_amd64_ctr_enc;
+ELF(.size _gcry_twofish_amd64_ctr_enc,.-_gcry_twofish_amd64_ctr_enc;)
 
 .align 8
 .globl _gcry_twofish_amd64_cbc_dec
-.type   _gcry_twofish_amd64_cbc_dec,@function;
+ELF(.type   _gcry_twofish_amd64_cbc_dec,@function;)
 _gcry_twofish_amd64_cbc_dec:
 	/* input:
 	 *	%rdi: ctx, CTX
@@ -659,11 +666,11 @@ _gcry_twofish_amd64_cbc_dec:
 	addq $(9 * 8), %rsp;
 
 	ret;
-.size _gcry_twofish_amd64_cbc_dec,.-_gcry_twofish_amd64_cbc_dec;
+ELF(.size _gcry_twofish_amd64_cbc_dec,.-_gcry_twofish_amd64_cbc_dec;)
 
 .align 8
 .globl _gcry_twofish_amd64_cfb_dec
-.type   _gcry_twofish_amd64_cfb_dec,@function;
+ELF(.type   _gcry_twofish_amd64_cfb_dec,@function;)
 _gcry_twofish_amd64_cfb_dec:
 	/* input:
 	 *	%rdi: ctx, CTX
@@ -725,7 +732,7 @@ _gcry_twofish_amd64_cfb_dec:
 	addq $(8 * 8), %rsp;
 
 	ret;
-.size _gcry_twofish_amd64_cfb_dec,.-_gcry_twofish_amd64_cfb_dec;
+ELF(.size _gcry_twofish_amd64_cfb_dec,.-_gcry_twofish_amd64_cfb_dec;)
 
 #endif /*USE_TWOFISH*/
 #endif /*__x86_64*/
diff --git a/cipher/twofish.c b/cipher/twofish.c
index ecd76e35..ce83fadf 100644
--- a/cipher/twofish.c
+++ b/cipher/twofish.c
@@ -53,7 +53,8 @@
 
 /* USE_AMD64_ASM indicates whether to use AMD64 assembly code. */
 #undef USE_AMD64_ASM
-#if defined(__x86_64__) && defined(HAVE_COMPATIBLE_GCC_AMD64_PLATFORM_AS)
+#if defined(__x86_64__) && (defined(HAVE_COMPATIBLE_GCC_AMD64_PLATFORM_AS) || \
+    defined(HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS))
 # define USE_AMD64_ASM 1
 #endif
 
@@ -754,6 +755,77 @@ extern void _gcry_twofish_amd64_cbc_dec(const TWOFISH_context *c, byte *out,
 extern void _gcry_twofish_amd64_cfb_dec(const TWOFISH_context *c, byte *out,
 					const byte *in, byte *iv);
 
+#ifdef HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS
+static inline void
+call_sysv_fn (const void *fn, const void *arg1, const void *arg2,
+              const void *arg3, const void *arg4)
+{
+  /* Call SystemV ABI function without storing non-volatile XMM registers,
+   * as target function does not use vector instruction sets. */
+  asm volatile ("callq *%0\n\t"
+                : "+a" (fn),
+                  "+D" (arg1),
+                  "+S" (arg2),
+                  "+d" (arg3),
+                  "+c" (arg4)
+                :
+                : "cc", "memory", "r8", "r9", "r10", "r11");
+}
+#endif
+
+static inline void
+twofish_amd64_encrypt_block(const TWOFISH_context *c, byte *out, const byte *in)
+{
+#ifdef HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS
+  call_sysv_fn(_gcry_twofish_amd64_encrypt_block, c, out, in, NULL);
+#else
+  _gcry_twofish_amd64_encrypt_block(c, out, in);
+#endif
+}
+
+static inline void
+twofish_amd64_decrypt_block(const TWOFISH_context *c, byte *out, const byte *in)
+{
+#ifdef HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS
+  call_sysv_fn(_gcry_twofish_amd64_decrypt_block, c, out, in, NULL);
+#else
+  _gcry_twofish_amd64_decrypt_block(c, out, in);
+#endif
+}
+
+static inline void
+twofish_amd64_ctr_enc(const TWOFISH_context *c, byte *out, const byte *in,
+                      byte *ctr)
+{
+#ifdef HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS
+  call_sysv_fn(_gcry_twofish_amd64_ctr_enc, c, out, in, ctr);
+#else
+  _gcry_twofish_amd64_ctr_enc(c, out, in, ctr);
+#endif
+}
+
+static inline void
+twofish_amd64_cbc_dec(const TWOFISH_context *c, byte *out, const byte *in,
+                      byte *iv)
+{
+#ifdef HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS
+  call_sysv_fn(_gcry_twofish_amd64_cbc_dec, c, out, in, iv);
+#else
+  _gcry_twofish_amd64_cbc_dec(c, out, in, iv);
+#endif
+}
+
+static inline void
+twofish_amd64_cfb_dec(const TWOFISH_context *c, byte *out, const byte *in,
+                      byte *iv)
+{
+#ifdef HAVE_COMPATIBLE_GCC_WIN64_PLATFORM_AS
+  call_sysv_fn(_gcry_twofish_amd64_cfb_dec, c, out, in, iv);
+#else
+  _gcry_twofish_amd64_cfb_dec(c, out, in, iv);
+#endif
+}
+
 #elif defined(USE_ARM_ASM)
 
 /* Assembly implementations of Twofish. */
@@ -833,7 +905,7 @@ static unsigned int
 twofish_encrypt (void *context, byte *out, const byte *in)
 {
   TWOFISH_context *ctx = context;
-  _gcry_twofish_amd64_encrypt_block(ctx, out, in);
+  twofish_amd64_encrypt_block(ctx, out, in);
   return /*burn_stack*/ (4*sizeof (void*));
 }
 
@@ -900,7 +972,7 @@ static unsigned int
 twofish_decrypt (void *context, byte *out, const byte *in)
 {
   TWOFISH_context *ctx = context;
-  _gcry_twofish_amd64_decrypt_block(ctx, out, in);
+  twofish_amd64_decrypt_block(ctx, out, in);
   return /*burn_stack*/ (4*sizeof (void*));
 }
 
@@ -980,7 +1052,7 @@ _gcry_twofish_ctr_enc(void *context, unsigned char *ctr, void *outbuf_arg,
     /* Process data in 3 block chunks. */
     while (nblocks >= 3)
       {
-        _gcry_twofish_amd64_ctr_enc(ctx, outbuf, inbuf, ctr);
+        twofish_amd64_ctr_enc(ctx, outbuf, inbuf, ctr);
 
         nblocks -= 3;
         outbuf += 3 * TWOFISH_BLOCKSIZE;
@@ -1038,7 +1110,7 @@ _gcry_twofish_cbc_dec(void *context, unsigned char *iv, void *outbuf_arg,
     /* Process data in 3 block chunks. */
     while (nblocks >= 3)
       {
-        _gcry_twofish_amd64_cbc_dec(ctx, outbuf, inbuf, iv);
+        twofish_amd64_cbc_dec(ctx, outbuf, inbuf, iv);
 
         nblocks -= 3;
         outbuf += 3 * TWOFISH_BLOCKSIZE;
@@ -1087,7 +1159,7 @@ _gcry_twofish_cfb_dec(void *context, unsigned char *iv, void *outbuf_arg,
     /* Process data in 3 block chunks. */
     while (nblocks >= 3)
       {
-        _gcry_twofish_amd64_cfb_dec(ctx, outbuf, inbuf, iv);
+        twofish_amd64_cfb_dec(ctx, outbuf, inbuf, iv);
 
         nblocks -= 3;
         outbuf += 3 * TWOFISH_BLOCKSIZE;