Fix cross-compilation for ARM FPU variants and add RISC-V support

2026-04-05 21:51:03 +00:00 · 2026-03-19 12:18:41 +01:00
parent a1cf319999
commit bd01e7d8a5
6 changed files with 62 additions and 44 deletions
--- a/3rdparty/blake3.cmake
+++ b/3rdparty/blake3.cmake
@@ -79,9 +79,22 @@ else()
        message(STATUS "BLAKE3: Enabling NEON for AArch64")
        target_sources(blake3 PRIVATE ${BLAKE3_C_DIRECTORY}/blake3_neon.c)
    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "arm")
-        message(STATUS "BLAKE3: Attempting NEON on 32-bit ARM")
-        target_sources(blake3 PRIVATE ${BLAKE3_C_DIRECTORY}/blake3_neon.c)
-        target_compile_options(blake3 PRIVATE -mfpu=neon)
+        # Test whether the target already supports NEON (via CMAKE_C_FLAGS or
+        # compiler defaults).  We must NOT add -mfpu=neon ourselves: GCC uses
+        # last-wins for -mfpu, so appending it would silently override a user's
+        # -mfpu=vfpv3-d16 and produce a binary that crashes on non-NEON hardware.
+        include(CheckCSourceCompiles)
+        check_c_source_compiles("
+            #include <arm_neon.h>
+            int main(void) { uint32x4_t v = vdupq_n_u32(0); return vgetq_lane_u32(v, 0); }
+        " BLAKE3_NEON_COMPILES)
+        if(BLAKE3_NEON_COMPILES)
+            message(STATUS "BLAKE3: Enabling NEON on 32-bit ARM")
+            target_sources(blake3 PRIVATE ${BLAKE3_C_DIRECTORY}/blake3_neon.c)
+            target_compile_definitions(blake3 PRIVATE BLAKE3_USE_NEON=1)
+        else()
+            message(STATUS "BLAKE3: 32-bit ARM without NEON -> portable only")
+        endif()
    else()
        message(STATUS "BLAKE3: Unknown arch -> portable only")
    endif()
--- a/3rdparty/xxhash.cmake
+++ b/3rdparty/xxhash.cmake
@@ -35,13 +35,16 @@ if(CMAKE_BUILD_TYPE STREQUAL "Release")
    else()
        target_compile_options(xxhash PRIVATE -O3 -ffast-math)

-        # Enable specific optimizations for x86/x64
-        if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64" OR
-           ${CMAKE_SYSTEM_PROCESSOR} MATCHES "AMD64" OR
-           ${CMAKE_SYSTEM_PROCESSOR} MATCHES "i386" OR
-           ${CMAKE_SYSTEM_PROCESSOR} MATCHES "i686")
-            if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
-                target_compile_options(xxhash PRIVATE -march=core2 -mtune=westmere)
+        # Enable specific optimizations for native x86/x64 builds.
+        # Cross-compilation toolchain files provide their own -march/-mtune.
+        if(NOT CMAKE_CROSSCOMPILING)
+            if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64" OR
+               ${CMAKE_SYSTEM_PROCESSOR} MATCHES "AMD64" OR
+               ${CMAKE_SYSTEM_PROCESSOR} MATCHES "i386" OR
+               ${CMAKE_SYSTEM_PROCESSOR} MATCHES "i686")
+                if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
+                    target_compile_options(xxhash PRIVATE -march=core2 -mtune=westmere)
+                endif()
            endif()
        endif()
    endif()
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -103,7 +103,8 @@ if("${CMAKE_SIZEOF_VOID_P}" MATCHES "8"
   OR "${CMAKE_C_COMPILER_ARCHITECTURE_ID}" MATCHES "x64"
   OR "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "x86_64"
   OR "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "AMD64"
-   OR "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "aarch64")
+   OR "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "aarch64"
+   OR "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "riscv64")
  set(ARCHITECTURE_IS_64BIT TRUE)
 endif()

@@ -118,33 +119,37 @@ if("${CMAKE_BUILD_TYPE}" MATCHES "Release")
  else()
    add_compile_options(-ffast-math -O3)

-    if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64"
-       OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "i686"
-       OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "AMD64")
-      if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
-        add_compile_options(-march=core2 -mtune=westmere -mfpmath=sse)
-      endif()
+    # Architecture-specific flags are only set for native builds.
+    # Cross-compilation toolchain files provide their own -march/-mtune/-mfpu.
+    if(NOT CMAKE_CROSSCOMPILING)
+      if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64"
+         OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "i686"
+         OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "AMD64")
+        if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
+          add_compile_options(-march=core2 -mtune=westmere -mfpmath=sse)
+        endif()

-      add_compile_options(-msse3)
+        add_compile_options(-msse3)

-      if(NOT "${CMAKE_C_PLATFORM_ID}" MATCHES "MinGW")
-        add_compile_options(-flto)
-      endif()
-    elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
-      if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
-        add_compile_options(-march=armv8-a)
-      endif()
+        if(NOT "${CMAKE_C_PLATFORM_ID}" MATCHES "MinGW")
+          add_compile_options(-flto)
+        endif()
+      elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+        if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
+          add_compile_options(-march=armv8-a)
+        endif()

-      if(NOT "${CMAKE_C_PLATFORM_ID}" MATCHES "MinGW")
-        add_compile_options(-flto)
-      endif()
-    elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7l" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
-      if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
-        add_compile_options(-march=armv7+fp -mfpu=vfpv3-d16)
-      endif()
-    elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "mips")
-      if(NOT "${CMAKE_C_PLATFORM_ID}" MATCHES "MinGW")
-        add_compile_options(-flto)
+        if(NOT "${CMAKE_C_PLATFORM_ID}" MATCHES "MinGW")
+          add_compile_options(-flto)
+        endif()
+      elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7l" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
+        if(NOT "${CMAKE_C_COMPILER_ID}" MATCHES "AppleClang")
+          add_compile_options(-march=armv7+fp -mfpu=vfpv3-d16)
+        endif()
+      elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "mips")
+        if(NOT "${CMAKE_C_PLATFORM_ID}" MATCHES "MinGW")
+          add_compile_options(-flto)
+        endif()
      endif()
    endif()
  endif()
--- a/include/aaruformat/simd.h
+++ b/include/aaruformat/simd.h
@@ -86,18 +86,14 @@

 #ifdef __clang__
 #define TARGET_ARMV8_WITH_CRC __attribute__((target("armv8-a,crc")))
+#define TARGET_WITH_CRYPTO    __attribute__((target("armv8-a,crypto")))
 #else
 #define TARGET_ARMV8_WITH_CRC __attribute__((target("arch=armv8-a+crc")))
+#define TARGET_WITH_CRYPTO    __attribute__((target("fpu=crypto-neon-fp-armv8")))
 #endif

 #endif  // __ARM_ARCH >= 8

-#ifdef __clang__
-#define TARGET_WITH_CRYPTO __attribute__((target("armv8-a,crypto")))
-#else
-#define TARGET_WITH_CRYPTO __attribute__((target("fpu=crypto-neon-fp-armv8")))
-#endif
-
 #ifdef __clang__
 #define TARGET_WITH_SIMD __attribute__((target("neon")))
 #else
--- a/src/crc64/arm_vmull.c
+++ b/src/crc64/arm_vmull.c
@@ -32,7 +32,7 @@

 #include "arm_vmull.h"

-#if !defined(__MINGW32__) && (!defined(__ANDROID__) || !defined(__arm__))
+#ifdef TARGET_WITH_CRYPTO
 TARGET_WITH_CRYPTO static uint64x2_t sse2neon_vmull_p64_crypto(uint64x1_t _a, uint64x1_t _b)
 {
    poly64_t a = vget_lane_p64(vreinterpret_p64_u64(_a), 0);
@@ -43,8 +43,7 @@ TARGET_WITH_CRYPTO static uint64x2_t sse2neon_vmull_p64_crypto(uint64x1_t _a, ui

 TARGET_WITH_SIMD uint64x2_t sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b)
 {
-#if !defined(__MINGW32__) && (!defined(__ANDROID__) || !defined(__arm__))
-    // Wraps vmull_p64
+#ifdef TARGET_WITH_CRYPTO
    if(have_arm_crypto()) return sse2neon_vmull_p64_crypto(_a, _b);
 #endif

--- a/src/crc64/arm_vmull.h
+++ b/src/crc64/arm_vmull.h
@@ -21,7 +21,9 @@

 #if defined(__aarch64__) || defined(_M_ARM64) || defined(__arm__) || defined(_M_ARM)

+#ifdef TARGET_WITH_CRYPTO
 TARGET_WITH_CRYPTO static uint64x2_t sse2neon_vmull_p64_crypto(uint64x1_t _a, uint64x1_t _b);
+#endif
 TARGET_WITH_SIMD uint64x2_t          sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b);
 TARGET_WITH_SIMD uint64x2_t          mm_shuffle_epi8(uint64x2_t a, uint64x2_t b);
 TARGET_WITH_SIMD uint64x2_t          mm_srli_si128(uint64x2_t a, int imm);