Merge pull request #8044 from daverodgman/msft-aarch64

Better support for MSVC aarch64 aka ARM64 and ARM64EC
2023-12-01 07:48:26 +00:00 · 2023-12-01 07:48:26 +00:00 · 422951b9ed
commit 422951b9ed
parent 10dfe76425 059f66ce7c
8 changed files with 36 additions and 40 deletions
--- a/include/mbedtls/build_info.h
+++ b/include/mbedtls/build_info.h
@ -62,6 +62,11 @@
 #define MBEDTLS_ARCH_IS_X86
 #endif

+#if !defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64) && \
+    (defined(_M_ARM64) || defined(_M_ARM64EC))
+#define MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64
+#endif
+
 /* This is defined if the architecture is Armv8-A, or higher */
 #if !defined(MBEDTLS_ARCH_IS_ARMV8_A)
 #if defined(__ARM_ARCH) && defined(__ARM_ARCH_PROFILE)
--- a/library/aesce.c
+++ b/library/aesce.c
@ -23,7 +23,7 @@
 * By defining the macros ourselves we gain access to those declarations without
 * requiring -march on the command line.
 *
- * `arm_neon.h` could be included by any header file, so we put these defines
+ * `arm_neon.h` is included by common.h, so we put these defines
 * at the top of this file, before any includes.
 */
 #define __ARM_FEATURE_CRYPTO 1
@ -76,12 +76,6 @@
 #    endif
 #endif

-#ifdef __ARM_NEON
-#include <arm_neon.h>
-#else
-#error "Target does not support NEON instructions"
-#endif
-
 #if !(defined(__ARM_FEATURE_CRYPTO) || defined(__ARM_FEATURE_AES)) || \
    defined(MBEDTLS_ENABLE_ARM_CRYPTO_EXTENSIONS_COMPILER_FLAG)
 #   if defined(__ARMCOMPILER_VERSION)
@ -619,6 +613,6 @@ void mbedtls_aesce_gcm_mult(unsigned char c[16],
 #undef MBEDTLS_POP_TARGET_PRAGMA
 #endif

-#endif /* MBEDTLS_ARCH_IS_ARMV8_A */
+#endif /* MBEDTLS_AESCE_HAVE_CODE */

 #endif /* MBEDTLS_AESCE_C */
--- a/library/aesce.h
+++ b/library/aesce.h
@ -20,7 +20,8 @@
 #include "mbedtls/aes.h"


-#if defined(MBEDTLS_AESCE_C) && defined(MBEDTLS_ARCH_IS_ARMV8_A) && defined(__ARM_NEON) \
+#if defined(MBEDTLS_AESCE_C) \
+    && defined(MBEDTLS_ARCH_IS_ARMV8_A) && defined(MBEDTLS_HAVE_NEON_INTRINSICS) \
    && (defined(MBEDTLS_COMPILER_IS_GCC) || defined(__clang__) || defined(MSC_VER))

 /* MBEDTLS_AESCE_HAVE_CODE is defined if we have a suitable target platform, and a
@ -129,7 +130,7 @@ int mbedtls_aesce_setkey_enc(unsigned char *rk,
 #error "AES hardware acceleration not supported on this platform / compiler"
 #endif

-#endif /* MBEDTLS_AESCE_C && MBEDTLS_ARCH_IS_ARMV8_A && __ARM_NEON &&
+#endif /* MBEDTLS_AESCE_C && MBEDTLS_ARCH_IS_ARMV8_A && MBEDTLS_HAVE_NEON_INTRINSICS &&
          (MBEDTLS_COMPILER_IS_GCC || __clang__ || MSC_VER) */

 #endif /* MBEDTLS_AESCE_H */
--- a/library/alignment.h
+++ b/library/alignment.h
@ -23,11 +23,16 @@
 * efficient when this is not defined.
 */
 #if defined(__ARM_FEATURE_UNALIGNED) \
-    || defined(__i386__) || defined(__amd64__) || defined(__x86_64__)
+    || defined(MBEDTLS_ARCH_IS_X86) || defined(MBEDTLS_ARCH_IS_X64) \
+    || defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64)
 /*
 * __ARM_FEATURE_UNALIGNED is defined where appropriate by armcc, gcc 7, clang 9
 * (and later versions) for Arm v7 and later; all x86 platforms should have
 * efficient unaligned access.
+ *
+ * https://learn.microsoft.com/en-us/cpp/build/arm64-windows-abi-conventions?view=msvc-170#alignment
+ * specifies that on Windows-on-Arm64, unaligned access is safe (except for uncached
+ * device memory).
 */
 #define MBEDTLS_EFFICIENT_UNALIGNED_ACCESS
 #endif
--- a/library/bignum.c
+++ b/library/bignum.c
@ -102,7 +102,8 @@ int mbedtls_mpi_lt_mpi_ct(const mbedtls_mpi *X,
 * about whether the assignment was made or not.
 * (Leaking information about the respective sizes of X and Y is ok however.)
 */
-#if defined(_MSC_VER) && defined(_M_ARM64) && (_MSC_FULL_VER < 193131103)
+#if defined(_MSC_VER) && defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64) && \
+    (_MSC_FULL_VER < 193131103)
 /*
 * MSVC miscompiles this function if it's inlined prior to Visual Studio 2022 version 17.1. See:
 * https://developercommunity.visualstudio.com/t/c-compiler-miscompiles-part-of-mbedtls-library-on/1646989
--- a/library/common.h
+++ b/library/common.h
@ -21,7 +21,11 @@

 #if defined(__ARM_NEON)
 #include <arm_neon.h>
-#endif /* __ARM_NEON */
+#define MBEDTLS_HAVE_NEON_INTRINSICS
+#elif defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64)
+#include <arm64_neon.h>
+#define MBEDTLS_HAVE_NEON_INTRINSICS
+#endif


 #if defined(__GNUC__) && !defined(__ARMCC_VERSION) && !defined(__clang__) \
@ -178,16 +182,16 @@ inline void mbedtls_xor(unsigned char *r, const unsigned char *a, const unsigned
 {
    size_t i = 0;
 #if defined(MBEDTLS_EFFICIENT_UNALIGNED_ACCESS)
-#if defined(__ARM_NEON) && \
-    (!defined(MBEDTLS_COMPILER_IS_GCC) || \
-    (defined(MBEDTLS_COMPILER_IS_GCC) && MBEDTLS_GCC_VERSION >= 70300))
+#if defined(MBEDTLS_HAVE_NEON_INTRINSICS) && \
+    (!(defined(MBEDTLS_COMPILER_IS_GCC) && MBEDTLS_GCC_VERSION < 70300))
+    /* Old GCC versions generate a warning here, so disable the NEON path for these compilers */
    for (; (i + 16) <= n; i += 16) {
        uint8x16_t v1 = vld1q_u8(a + i);
        uint8x16_t v2 = vld1q_u8(b + i);
        uint8x16_t x = veorq_u8(v1, v2);
        vst1q_u8(r + i, x);
    }
-#elif defined(__amd64__) || defined(__x86_64__) || defined(__aarch64__)
+#elif defined(MBEDTLS_ARCH_IS_X64) || defined(MBEDTLS_ARCH_IS_ARM64)
    /* This codepath probably only makes sense on architectures with 64-bit registers */
    for (; (i + 8) <= n; i += 8) {
        uint64_t x = mbedtls_get_unaligned_uint64(a + i) ^ mbedtls_get_unaligned_uint64(b + i);
@ -226,7 +230,7 @@ static inline void mbedtls_xor_no_simd(unsigned char *r,
 {
    size_t i = 0;
 #if defined(MBEDTLS_EFFICIENT_UNALIGNED_ACCESS)
-#if defined(__amd64__) || defined(__x86_64__) || defined(__aarch64__)
+#if defined(MBEDTLS_ARCH_IS_X64) || defined(MBEDTLS_ARCH_IS_ARM64)
    /* This codepath probably only makes sense on architectures with 64-bit registers */
    for (; (i + 8) <= n; i += 8) {
        uint64_t x = mbedtls_get_unaligned_uint64(a + i) ^ mbedtls_get_unaligned_uint64(b + i);
--- a/library/sha256.c
+++ b/library/sha256.c
@ -28,7 +28,7 @@
 * By defining the macros ourselves we gain access to those declarations without
 * requiring -march on the command line.
 *
- * `arm_neon.h` could be included by any header file, so we put these defines
+ * `arm_neon.h` is included by common.h, so we put these defines
 * at the top of this file, before any includes.
 */
 #define __ARM_FEATURE_CRYPTO 1
@ -62,9 +62,7 @@

 #  if defined(MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_IF_PRESENT) || \
    defined(MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_ONLY)
-#       ifdef __ARM_NEON
-#           include <arm_neon.h>
-#       else
+#       if !defined(MBEDTLS_HAVE_NEON_INTRINSICS)
 #           if defined(MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_IF_PRESENT)
 #               warning "Target does not support NEON instructions"
 #               undef MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_IF_PRESENT
@ -126,12 +124,7 @@
 #      include <signal.h>
 #    endif
 #  endif
-#elif defined(_M_ARM64)
-#  if defined(MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_IF_PRESENT) || \
-    defined(MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_ONLY)
-#    include <arm64_neon.h>
-#  endif
-#else
+#elif !defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64)
 #  undef MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_ONLY
 #  undef MBEDTLS_SHA256_USE_ARMV8_A_CRYPTO_IF_PRESENT
 #endif
@ -156,7 +149,7 @@ static int mbedtls_a64_crypto_sha256_determine_support(void)
 {
    return 1;
 }
-#elif defined(_M_ARM64)
+#elif defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64)
 #define WIN32_LEAN_AND_MEAN
 #include <Windows.h>
 #include <processthreadsapi.h>
--- a/library/sha512.c
+++ b/library/sha512.c
@ -19,7 +19,7 @@
 * By defining the macros ourselves we gain access to those declarations without
 * requiring -march on the command line.
 *
- * `arm_neon.h` could be included by any header file, so we put these defines
+ * `arm_neon.h` is included by common.h, so we put these defines
 * at the top of this file, before any includes.
 */
 #define __ARM_FEATURE_SHA512 1
@ -48,9 +48,7 @@
 #  if defined(MBEDTLS_SHA512_USE_A64_CRYPTO_IF_PRESENT) || \
    defined(MBEDTLS_SHA512_USE_A64_CRYPTO_ONLY)
 /* *INDENT-OFF* */
-#   ifdef __ARM_NEON
-#       include <arm_neon.h>
-#   else
+#   if !defined(MBEDTLS_HAVE_NEON_INTRINSICS)
 #       error "Target does not support NEON instructions"
 #   endif
 /*
@ -109,12 +107,7 @@
 #      include <signal.h>
 #    endif
 #  endif
-#elif defined(_M_ARM64)
-#  if defined(MBEDTLS_SHA512_USE_A64_CRYPTO_IF_PRESENT) || \
-    defined(MBEDTLS_SHA512_USE_A64_CRYPTO_ONLY)
-#    include <arm64_neon.h>
-#  endif
-#else
+#elif !defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64)
 #  undef MBEDTLS_SHA512_USE_A64_CRYPTO_ONLY
 #  undef MBEDTLS_SHA512_USE_A64_CRYPTO_IF_PRESENT
 #endif
@ -142,7 +135,7 @@ static int mbedtls_a64_crypto_sha512_determine_support(void)
                           NULL, 0);
    return ret == 0 && value != 0;
 }
-#elif defined(_M_ARM64)
+#elif defined(MBEDTLS_PLATFORM_IS_WINDOWS_ON_ARM64)
 /*
 * As of March 2022, there don't appear to be any PF_ARM_V8_* flags
 * available to pass to IsProcessorFeaturePresent() to check for