gcc+neon

Author: moon-chilled

jsrc/avxintrin-neon.h

@@ -2564,8 +2564,8 @@ FORCE_INLINE __m256 _mm256_castpd_ps(__m256d a)
 {
     __m256 b;
 
-    b.vect_f32[0] = vreinterpretq_f64_f32(a.vect_f64[0]);
-    b.vect_f32[1] = vreinterpretq_f64_f32(a.vect_f64[1]);
+    b.vect_f32[0] = vreinterpretq_f32_f64(a.vect_f64[0]);
+    b.vect_f32[1] = vreinterpretq_f32_f64(a.vect_f64[1]);
 
     return b;
 }
@@ -2574,8 +2574,8 @@ FORCE_INLINE __m256i _mm256_castpd_si256(__m256d a)
 {
     __m256i b;
 
-    b.vect_s64[0] = vreinterpretq_f64_s64(a.vect_f64[0]);
-    b.vect_s64[1] = vreinterpretq_f64_s64(a.vect_f64[1]);
+    b.vect_s64[0] = vreinterpretq_s64_f64(a.vect_f64[0]);
+    b.vect_s64[1] = vreinterpretq_s64_f64(a.vect_f64[1]);
 
     return b;
 }
@@ -2584,8 +2584,8 @@ FORCE_INLINE __m256d _mm256_castps_pd(__m256 a)
 {
     __m256d b;
 
-    b.vect_f64[0] = vreinterpretq_f32_f64(a.vect_f32[0]);
-    b.vect_f64[1] = vreinterpretq_f32_f64(a.vect_f32[1]);
+    b.vect_f64[0] = vreinterpretq_f64_f32(a.vect_f32[0]);
+    b.vect_f64[1] = vreinterpretq_f64_f32(a.vect_f32[1]);
 
     return b;
 }

jsrc/j.h

@@ -1098,8 +1098,8 @@ if(likely(!((I)jtinplace&JTWILLBEOPENED)))z=EPILOGNORET(z); RETF(z); \
 // AVX512  u=_mm256_cvtepi64_pd(_mm256_castpd_si256(u));
 #if 1
 #if defined(__aarch64__)
-#define  CVTEPI64(z,u)   z.vect_f64[0] = vcvtq_f64_s64(vreinterpretq_f64_s64(u.vect_f64[0])); \
-                         z.vect_f64[1] = vcvtq_f64_s64(vreinterpretq_f64_s64(u.vect_f64[1]));
+#define  CVTEPI64(z,u)   z.vect_f64[0] = vcvtq_f64_s64(vreinterpretq_s64_f64(u.vect_f64[0])); \
+                         z.vect_f64[1] = vcvtq_f64_s64(vreinterpretq_s64_f64(u.vect_f64[1]));
 #else
 
 #define  CVTEPI64(z,u) { __m256i u_lo = _mm256_castps_si256(_mm256_blend_ps(_mm256_castsi256_ps(magic_i_lo),_mm256_castpd_ps(u),0b01010101));         /* Blend the 32 lowest significant bits of u with magic_int_lo */ \

make2/build_libj.sh

@@ -92,7 +92,7 @@ fi
 
 if [ -z "${compiler##*gcc*}" ] || [ -z "${CC##*gcc*}" ]; then
 # gcc
-common="$OPENMP -fPIC -O2 -falign-functions=4 -fvisibility=hidden -fno-strict-aliasing -fwrapv -fno-stack-protector  \
+common="$OPENMP -fPIC -O2 -falign-functions=4 -fvisibility=hidden -fno-strict-aliasing -fwrapv -fno-stack-protector -flax-vector-conversions \
  -Werror -Wextra -Wno-unknown-warning-option \
  -Wno-cast-function-type \
  -Wno-clobbered \