Support Microsoft CL.EXE compiler

config/cfg.ml

@@ -1,30 +1,50 @@
-let std_flags = ["--std=c11"; "-Wall"; "-Wextra"; "-Wpedantic"; "-O3"]
+let std_flags = ["-Wall"]
 
 let () =
   let c = Configurator.V1.create "mirage-crypto" in
+  let ccomp_type_opt = Configurator.V1.ocaml_config_var c "ccomp_type" in
   let arch =
     let defines =
       Configurator.V1.C_define.import
         c
         ~includes:[]
-        [("__x86_64__", Switch); ("__i386__", Switch)]
+        [("__x86_64__", Switch); ("__i386__", Switch); ("_WIN64", Switch); ("_WIN32", Switch)]
     in
     match defines with
     | (_, Switch true) :: _ -> `x86_64
     | _ :: (_, Switch true) :: _ -> `x86
+    | _ :: _ :: (_, Switch true) :: _ -> `x86_64
+    | _ :: _ :: _ :: (_, Switch true) :: _ -> `x86
     | _ -> `unknown
   in
   let accelerate_flags =
-    match arch with
-    | `x86_64 -> [ "-DACCELERATE"; "-mssse3"; "-maes"; "-mpclmul" ]
+    match arch, ccomp_type_opt with
+    | `x86_64, Some ccomp_type when ccomp_type = "msvc" -> [ "-DACCELERATE" ]
+    | `x86_64, _ -> [ "-DACCELERATE"; "-mssse3"; "-maes"; "-mpclmul" ]
     | _ -> []
   in
   let ent_flags =
-    match arch with
-    | `x86_64 | `x86 -> [ "-DENTROPY"; "-mrdrnd"; "-mrdseed" ]
+    match arch, ccomp_type_opt with
+    | (`x86_64 | `x86), Some ccomp_type when ccomp_type = "msvc" -> [ "-DENTROPY" ]
+    | (`x86_64 | `x86), _ -> [ "-DENTROPY"; "-mrdrnd"; "-mrdseed" ]
     | _ -> []
   in
-  let flags = std_flags @ ent_flags in
+  let lang_flags =
+    match ccomp_type_opt with
+    | Some ccomp_type when ccomp_type = "msvc" -> ["/std:c11"]
+    | _ -> ["--std=c11"]
+  in
+  let warn_flags =
+    match ccomp_type_opt with
+    | Some ccomp_type when ccomp_type = "msvc" -> []
+    | _ -> ["-Wextra"; "-Wpedantic"]
+  in
+  let optimization_flags =
+    match ccomp_type_opt with
+    | Some ccomp_type when ccomp_type = "msvc" -> ["-O2"]
+    | _ -> ["-O3"]
+  in
+  let flags = std_flags @ ent_flags @ lang_flags @ warn_flags @ optimization_flags in
   let opt_flags = flags @ accelerate_flags in
   Configurator.V1.Flags.write_sexp "cflags_optimized.sexp" opt_flags;
   Configurator.V1.Flags.write_sexp "cflags.sexp" flags

ec-freestanding/Makefile

@@ -2,7 +2,7 @@ ifneq (, $(shell command -v opam))
 PKG_CONFIG_PATH ?= $(shell opam var prefix)/lib/pkgconfig
 endif
 
-EXISTS := $(shell PKG_CONFIG_PATH=$(PKG_CONFIG_PATH) pkg-config --exists ocaml-freestanding; echo $$?)
+EXISTS := $(shell PKG_CONFIG_PATH="$(PKG_CONFIG_PATH)" pkg-config --exists ocaml-freestanding; echo $$?)
 
 .PHONY: all clean
 all: libmirage_crypto_ec_freestanding_stubs.a
@@ -12,7 +12,7 @@ libmirage_crypto_ec_freestanding_stubs.a:
 	touch $@
 else
 CC ?= cc
-FREESTANDING_CFLAGS := $(shell PKG_CONFIG_PATH=$(PKG_CONFIG_PATH) pkg-config --cflags ocaml-freestanding)
+FREESTANDING_CFLAGS := $(shell PKG_CONFIG_PATH="$(PKG_CONFIG_PATH)" pkg-config --cflags ocaml-freestanding)
 DISCOVER_CFLAGS := $(shell sed 's/^(\(.*\))$$/\1/' ../ec/cflags_optimized.sexp | tr -d '"')
 CFLAGS := -DNDEBUG -O3 -I../ec/native -I../src/native $(DISCOVER_CFLAGS) $(FREESTANDING_CFLAGS)
 

src-freestanding/Makefile

@@ -2,7 +2,7 @@ ifneq (, $(shell command -v opam))
 PKG_CONFIG_PATH ?= $(shell opam var prefix)/lib/pkgconfig
 endif
 
-EXISTS := $(shell PKG_CONFIG_PATH=$(PKG_CONFIG_PATH) pkg-config --exists ocaml-freestanding; echo $$?)
+EXISTS := $(shell PKG_CONFIG_PATH="$(PKG_CONFIG_PATH)" pkg-config --exists ocaml-freestanding; echo $$?)
 
 .PHONY: all clean
 all: libmirage_crypto_freestanding_stubs.a
@@ -12,7 +12,7 @@ libmirage_crypto_freestanding_stubs.a:
 	touch $@
 else
 CC ?= cc
-FREESTANDING_CFLAGS := $(shell PKG_CONFIG_PATH=$(PKG_CONFIG_PATH) pkg-config --cflags ocaml-freestanding)
+FREESTANDING_CFLAGS := $(shell PKG_CONFIG_PATH="$(PKG_CONFIG_PATH)" pkg-config --cflags ocaml-freestanding)
 MIRAGE_CRYPTO_OPT_CFLAGS := $(shell sed 's/^(\(.*\))$$/\1/' ../src/cflags_optimized.sexp | tr -d '"')
 CFLAGS := -I../src/native $(MIRAGE_CRYPTO_OPT_CFLAGS) $(FREESTANDING_CFLAGS)
 

src/native/aes_aesni.c

@@ -27,7 +27,7 @@ static int _mc_aesni_rk_size (uint8_t rounds) {
   return (rounds + 1) * 16 + 15;
 }
 
-#if defined(__x86_64__)
+#if defined(__x86_64__) || defined(_WIN64)
 static inline __m128i* __rk (const void *rk) {
   return (__m128i *) (((uint64_t)rk + 15) & -16);
 }
@@ -48,10 +48,17 @@ static inline __m128i __mix (__m128i r1, __m128i r2) {
 
 #define __assist(r1, r2, mode) (__mix (r1, _mm_shuffle_epi32 (r2, mode)))
 
+#ifdef _MSC_VER
+static inline void __pack (__m128i *o1, __m128i *o2, __m128i r1, __m128i r2, __m128i r3) {
+  *o1 = _mm_castpd_si128 (_mm_shuffle_pd (_mm_castsi128_pd (r1), _mm_castsi128_pd (r2), 0));
+  *o2 = _mm_castpd_si128 (_mm_shuffle_pd (_mm_castsi128_pd (r2), _mm_castsi128_pd (r3), 1));
+}
+#else
 static inline void __pack (__m128i *o1, __m128i *o2, __m128i r1, __m128i r2, __m128i r3) {
   *o1 = (__m128i) _mm_shuffle_pd ((__m128d) r1, (__m128d) r2, 0);
   *o2 = (__m128i) _mm_shuffle_pd ((__m128d) r2, (__m128d) r3, 1);
 }
+#endif
 
 static inline void _mc_aesni_derive_e_key (const uint8_t *key, uint8_t *rk0, uint8_t rounds) {
 

src/native/bitfn.h

@@ -121,8 +121,8 @@ static inline void array_copy64(uint64_t *d, uint64_t *s, uint32_t nb)
 	while (nb--) *d++ = *s++;
 }
 
-#ifdef __BYTE_ORDER__
-#if __ORDER_LITTLE_ENDIAN__ == __BYTE_ORDER__
+#if defined(_MSC_VER) || defined(__BYTE_ORDER__)
+#if defined(_MSC_VER) || (__ORDER_LITTLE_ENDIAN__ == __BYTE_ORDER__)
 
 # define be32_to_cpu(a) bitfn_swap32(a)
 # define cpu_to_be32(a) bitfn_swap32(a)

src/native/detect_cpu_features.c

@@ -2,10 +2,53 @@
 
 #ifdef __mc_detect_features__
 
-#include <cpuid.h>
+#ifndef _MSC_VER
+# include <cpuid.h>
+#endif
 
 struct _mc_cpu_features mc_detected_cpu_features = { 0 };
 
+#ifdef _MSC_VER
+#define bit_PCLMUL ((int)1 << 1)
+#define bit_SSSE3 ((int)1 << 9)
+#define bit_AES ((int)1 << 25)
+#define bit_RDRND ((int)1 << 30)
+#define bit_RDSEED ((int)1 << 18)
+
+CAMLprim value
+mc_detect_cpu_features (__unit ()) {
+  int cpuInfo[4] = {-1};
+  int ebx;
+  int ecx;
+
+  __cpuid(cpuInfo, 0x00000000);
+  int max = cpuInfo[0];
+  if (max < 1) return Val_unit;
+
+  __cpuid(cpuInfo, 0x00000001);
+  ecx = cpuInfo[2];
+
+  if (ecx & bit_PCLMUL)
+    mc_detected_cpu_features.pclmul = 1;
+  if (ecx & bit_SSSE3)
+    mc_detected_cpu_features.ssse3 = 1;
+  if (ecx & bit_AES)
+    mc_detected_cpu_features.aesni = 1;
+  if (ecx & bit_RDRND)
+    mc_detected_cpu_features.rdrand = 1;
+
+  if (max > 7) {
+    __cpuid(cpuInfo, 0x00000007);
+    ebx = cpuInfo[1];
+    if (ebx & bit_RDSEED)
+      mc_detected_cpu_features.rdseed = 1;
+  }
+
+  return Val_unit;
+}
+
+#else
+
 CAMLprim value
 mc_detect_cpu_features (__unit ()) {
   unsigned int sig = 0, eax = 0, ebx = 0, ecx = 0, edx = 0;
@@ -32,6 +75,7 @@ mc_detect_cpu_features (__unit ()) {
 
   return Val_unit;
 }
+#endif /* _MSC_VER */
 
 #else /* __mc_detect_features__ */
 

src/native/entropy_cpu_stubs.c

@@ -22,6 +22,22 @@
 #endif
 #endif /* __i386__ || __x86_64__ */
 
+#if defined (_MSC_VER)
+#include <immintrin.h>
+
+#if defined (_WIN64)
+#define random_t unsigned long long
+#define _rdseed_step _rdseed64_step
+#define _rdrand_step _rdrand64_step
+
+#elif defined (_WIN32)
+#define random_t unsigned int
+#define _rdseed_step _rdseed32_step
+#define _rdrand_step _rdrand32_step
+#endif
+
+#endif /* _MSC_VER */
+
 #if defined (__arm__)
 /*
  * The ideal timing source on ARM are the performance counters, but these are
@@ -119,7 +135,7 @@ static inline uint64_t getticks(void)
 
 
 CAMLprim value mc_cycle_counter (value __unused(unit)) {
-#if defined (__i386__) || defined (__x86_64__)
+#if defined (__i386__) || defined (__x86_64__) || defined (_MSC_VER)
   return Val_long (__rdtsc ());
 #elif defined (__arm__) || defined (__aarch64__)
   return Val_long (read_virtual_count ());

src/native/ghash_ctmul.c

@@ -39,7 +39,7 @@
 #include "mirage_crypto.h"
 #include <string.h>
 
-#if defined (__i386__) || defined (__arm__)
+#if defined (__i386__) || defined (__arm__) || defined(_MSC_VER)
 
 /*
  * We cannot really autodetect whether multiplications are "slow" or

src/native/ghash_generic.c

@@ -16,7 +16,11 @@
  * !LARGE_TABLES -> 8K per key, ~3x slower. */
 #define __MC_GHASH_LARGE_TABLES
 
-#ifdef ARCH_64BIT
+/* 64-bit Windows sets ARCH_64BIT but 128-bit integers are not supported
+ * by the Microsoft compiler. Drop down to 32-bit for MSVC;
+ * ghash_ctmul.c will implement ghash for MSVC.
+ */
+#if defined(ARCH_64BIT) && !defined(_MSC_VER)
 
 #define __set_uint128_t(w1, w0) (((__uint128_t) w1 << 64) | w0)
 

src/native/mirage_crypto.h

@@ -9,7 +9,11 @@
 #include <caml/bigarray.h>
 
 #ifdef ACCELERATE
-#include <x86intrin.h>
+# ifdef _MSC_VER
+#  include <intrin.h>
+# else
+#  include <x86intrin.h>
+# endif
 #define __mc_ACCELERATE__
 #define __mc_detect_features__
 #endif
@@ -47,16 +51,23 @@ extern struct _mc_cpu_features mc_detected_cpu_features;
 
 #endif /* __mc_ACCELERATE__ */
 
-#if defined (__x86_64__) || defined (__aarch64__) || defined (__powerpc64__) || (64 == __riscv_xlen) || defined (__s390x__)
+#if defined (__x86_64__) || defined (__aarch64__) || defined (__powerpc64__) || (64 == __riscv_xlen) || defined (__s390x__) || (1 == _WIN64)
 #define ARCH_64BIT
-#elif defined (__i386__) || defined (__arm__) || (32 == __riscv_xlen)
+#elif defined (__i386__) || defined (__arm__) || (32 == __riscv_xlen) || (1 == _WIN32)
 #define ARCH_32BIT
 #else
 #error "unsupported platform"
 #endif
 
 #ifndef __unused
-#define __unused(x) x __attribute__((unused))
+# if defined(_MSC_VER) && _MSC_VER >= 1500
+#  define __unused(x) __pragma( warning (push) ) \
+    __pragma( warning (disable:4189 ) ) \
+    x \
+    __pragma( warning (pop))
+# else
+#  define __unused(x) x __attribute__((unused))
+# endif
 #endif
 #define __unit() value __unused(_)
 

src/native/poly1305-donna.c

@@ -7,7 +7,10 @@ typedef struct poly1305_context {
         unsigned char opaque[136];
 } poly1305_context;
 
-#ifdef ARCH_64BIT
+/* 64-bit Windows sets ARCH_64BIT but poly1305-donna-64 requires 128-bit integers
+ * that are not supported by the Microsoft compiler. Drop down to 32-bit for MSVC.
+ */
+#if defined(ARCH_64BIT) && !defined(_MSC_VER)
 #include "poly1305-donna-64.h"
 #else
 #include "poly1305-donna-32.h"