f.write("""
#include <folly/Portability.h>
-#if (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2)
#include <stdint.h>
-#include <x86intrin.h>
namespace folly {
namespace detail {
-extern const __m128i groupVarintSSEMasks[] = {
+#if (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2)
+alignas(16) extern const uint64_t groupVarintSSEMasks[512] = {
""")
# Compute SSE masks
# 0xff: set corresponding byte in result to 0
for k in range(d, 4):
vals[j] |= 0xff << (8 * k)
- f.write(" {{static_cast<int64_t>(0x{1:08x}{0:08x}), "
- "static_cast<int64_t>(0x{3:08x}{2:08x})}},\n".format(*vals))
+ f.write(" 0x{1:08x}{0:08x}ULL, "
+ "0x{3:08x}{2:08x}ULL,\n".format(*vals))
f.write("};\n"
- "\n"
- "extern const uint8_t groupVarintLengths[] = {\n")
+ "#endif /*#if (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2)*/\n"
+ "\n"
+ "extern const uint8_t groupVarintLengths[] = {\n")
# Also compute total encoded lengths, including key byte
for i in range(0, 256):
f.write("""
};
-} // namespace detail
-} // namespace folly
-#endif /* (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2) */
+} // namespace detail
+} // namespace folly
""")
def main():