Add LoongArch LSX and LASX support (#2159)

* Add LoongArch SX support

* Add LoongArch ASX support

CMakeLists.txt

@@ -106,6 +106,24 @@ if(
   )
 endif()
 
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(loongarch64)$")
+    option(SIMDJSON_PREFER_LSX "Prefer LoongArch SX" ON)
+    include(CheckCXXCompilerFlag)
+    check_cxx_compiler_flag(-mlasx COMPILER_SUPPORTS_LASX)
+    check_cxx_compiler_flag(-mlsx COMPILER_SUPPORTS_LSX)
+    if(COMPILER_SUPPORTS_LASX AND NOT SIMDJSON_PREFER_LSX)
+        simdjson_add_props(
+            target_compile_options PRIVATE
+            -mlasx
+        )
+    elseif(COMPILER_SUPPORTS_LSX)
+        simdjson_add_props(
+            target_compile_options PRIVATE
+            -mlsx
+        )
+    endif()
+endif()
+
 # GCC and Clang have horrendous Debug builds when using SIMD.
 # A common fix is to use '-Og' instead.
 # bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412

include/simdjson/builtin.h

@@ -20,6 +20,10 @@
 #include "simdjson/ppc64.h"
 #elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(westmere)
 #include "simdjson/westmere.h"
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lsx)
+#include "simdjson/lsx.h"
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lasx)
+#include "simdjson/lasx.h"
 #else
 #error Unknown SIMDJSON_BUILTIN_IMPLEMENTATION
 #endif

include/simdjson/builtin/base.h

@@ -17,6 +17,10 @@ namespace simdjson {
   namespace ppc64 {}
 #elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(westmere)
   namespace westmere {}
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lsx)
+  namespace lsx {}
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lasx)
+  namespace lasx {}
 #else
 #error Unknown SIMDJSON_BUILTIN_IMPLEMENTATION
 #endif

include/simdjson/builtin/implementation.h

@@ -19,6 +19,10 @@
 #include "simdjson/ppc64/implementation.h"
 #elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(westmere)
 #include "simdjson/westmere/implementation.h"
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lsx)
+#include "simdjson/lsx/implementation.h"
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lasx)
+#include "simdjson/lasx/implementation.h"
 #else
 #error Unknown SIMDJSON_BUILTIN_IMPLEMENTATION
 #endif

include/simdjson/builtin/ondemand.h

@@ -20,6 +20,10 @@
 #include "simdjson/ppc64/ondemand.h"
 #elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(westmere)
 #include "simdjson/westmere/ondemand.h"
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lsx)
+#include "simdjson/lsx/ondemand.h"
+#elif SIMDJSON_BUILTIN_IMPLEMENTATION_IS(lasx)
+#include "simdjson/lasx/ondemand.h"
 #else
 #error Unknown SIMDJSON_BUILTIN_IMPLEMENTATION
 #endif

include/simdjson/generic/base.h

@@ -17,6 +17,10 @@
 #include "simdjson/arm64/begin.h"
 #elif SIMDJSON_IMPLEMENTATION_PPC64
 #include "simdjson/ppc64/begin.h"
+#elif SIMDJSON_IMPLEMENTATION_LSX
+#include "simdjson/lsx/begin.h"
+#elif SIMDJSON_IMPLEMENTATION_LASX
+#include "simdjson/lasx/begin.h"
 #elif SIMDJSON_IMPLEMENTATION_FALLBACK
 #include "simdjson/fallback/begin.h"
 #else

include/simdjson/implementation_detection.h

@@ -10,6 +10,8 @@
 #define SIMDJSON_IMPLEMENTATION_ID_icelake 4
 #define SIMDJSON_IMPLEMENTATION_ID_ppc64 5
 #define SIMDJSON_IMPLEMENTATION_ID_westmere 6
+#define SIMDJSON_IMPLEMENTATION_ID_lsx 7
+#define SIMDJSON_IMPLEMENTATION_ID_lasx 8
 
 #define SIMDJSON_IMPLEMENTATION_ID_FOR(IMPL) SIMDJSON_CAT(SIMDJSON_IMPLEMENTATION_ID_, IMPL)
 #define SIMDJSON_IMPLEMENTATION_ID SIMDJSON_IMPLEMENTATION_ID_FOR(SIMDJSON_IMPLEMENTATION)
@@ -74,9 +76,23 @@
 #endif
 #define SIMDJSON_CAN_ALWAYS_RUN_PPC64 SIMDJSON_IMPLEMENTATION_PPC64 && SIMDJSON_IS_PPC64 && SIMDJSON_IS_PPC64_VMX
 
+#ifndef SIMDJSON_IMPLEMENTATION_LASX
+#define SIMDJSON_IMPLEMENTATION_LASX (SIMDJSON_IS_LOONGARCH64 && __loongarch_asx)
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_LASX (SIMDJSON_IMPLEMENTATION_LASX)
+
+#ifndef SIMDJSON_IMPLEMENTATION_LSX
+#if SIMDJSON_CAN_ALWAYS_RUN_LASX
+#define SIMDJSON_IMPLEMENTATION_LSX 0
+#else
+#define SIMDJSON_IMPLEMENTATION_LSX (SIMDJSON_IS_LOONGARCH64 && __loongarch_sx)
+#endif
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_LSX (SIMDJSON_IMPLEMENTATION_LSX)
+
 // Default Fallback to on unless a builtin implementation has already been selected.
 #ifndef SIMDJSON_IMPLEMENTATION_FALLBACK
-#if SIMDJSON_CAN_ALWAYS_RUN_ARM64 || SIMDJSON_CAN_ALWAYS_RUN_ICELAKE || SIMDJSON_CAN_ALWAYS_RUN_HASWELL || SIMDJSON_CAN_ALWAYS_RUN_WESTMERE || SIMDJSON_CAN_ALWAYS_RUN_PPC64
+#if SIMDJSON_CAN_ALWAYS_RUN_ARM64 || SIMDJSON_CAN_ALWAYS_RUN_ICELAKE || SIMDJSON_CAN_ALWAYS_RUN_HASWELL || SIMDJSON_CAN_ALWAYS_RUN_WESTMERE || SIMDJSON_CAN_ALWAYS_RUN_PPC64 || SIMDJSON_CAN_ALWAYS_RUN_LSX || SIMDJSON_CAN_ALWAYS_RUN_LASX
 // if anything at all except fallback can always run, then disable fallback.
 #define SIMDJSON_IMPLEMENTATION_FALLBACK 0
 #else
@@ -98,6 +114,10 @@
 #define SIMDJSON_BUILTIN_IMPLEMENTATION arm64
 #elif SIMDJSON_CAN_ALWAYS_RUN_PPC64
 #define SIMDJSON_BUILTIN_IMPLEMENTATION ppc64
+#elif SIMDJSON_CAN_ALWAYS_RUN_LSX
+#define SIMDJSON_BUILTIN_IMPLEMENTATION lsx
+#elif SIMDJSON_CAN_ALWAYS_RUN_LASX
+#define SIMDJSON_BUILTIN_IMPLEMENTATION lasx
 #elif SIMDJSON_CAN_ALWAYS_RUN_FALLBACK
 #define SIMDJSON_BUILTIN_IMPLEMENTATION fallback
 #else

include/simdjson/internal/instruction_set.h

@@ -66,7 +66,9 @@ enum instruction_set {
   AVX512CD = 0x2000,
   AVX512BW = 0x4000,
   AVX512VL = 0x8000,
-  AVX512VBMI2 = 0x10000
+  AVX512VBMI2 = 0x10000,
+  LSX = 0x20000,
+  LASX = 0x40000,
 };
 
 } // namespace internal

include/simdjson/lasx.h

@@ -0,0 +1,8 @@
+#ifndef SIMDJSON_LASX_H
+#define SIMDJSON_LASX_H
+
+#include "simdjson/lasx/begin.h"
+#include "simdjson/generic/amalgamated.h"
+#include "simdjson/lasx/end.h"
+
+#endif // SIMDJSON_LASX_H

include/simdjson/lasx/base.h

@@ -0,0 +1,26 @@
+#ifndef SIMDJSON_LASX_BASE_H
+#define SIMDJSON_LASX_BASE_H
+
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/base.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+namespace simdjson {
+/**
+ * Implementation for LASX.
+ */
+namespace lasx {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_BASE_H

include/simdjson/lasx/begin.h

@@ -0,0 +1,10 @@
+#define SIMDJSON_IMPLEMENTATION lasx
+#include "simdjson/lasx/base.h"
+#include "simdjson/lasx/intrinsics.h"
+#include "simdjson/lasx/bitmanipulation.h"
+#include "simdjson/lasx/bitmask.h"
+#include "simdjson/lasx/numberparsing_defs.h"
+#include "simdjson/lasx/simd.h"
+#include "simdjson/lasx/stringparsing_defs.h"
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1

include/simdjson/lasx/bitmanipulation.h

@@ -0,0 +1,50 @@
+#ifndef SIMDJSON_LASX_BITMANIPULATION_H
+#define SIMDJSON_LASX_BITMANIPULATION_H
+
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/lasx/base.h"
+#include "simdjson/lasx/intrinsics.h"
+#include "simdjson/lasx/bitmask.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __lasx_xvpickve2gr_w(__lasx_xvpcnt_d(__m256i(v4u64{input_num, 0, 0, 0})), 0);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_BITMANIPULATION_H

include/simdjson/lasx/bitmask.h

@@ -0,0 +1,31 @@
+#ifndef SIMDJSON_LASX_BITMASK_H
+#define SIMDJSON_LASX_BITMASK_H
+
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/lasx/base.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif

include/simdjson/lasx/end.h

@@ -0,0 +1,6 @@
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/lasx/base.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+#undef SIMDJSON_IMPLEMENTATION

include/simdjson/lasx/implementation.h

@@ -0,0 +1,31 @@
+#ifndef SIMDJSON_LASX_IMPLEMENTATION_H
+#define SIMDJSON_LASX_IMPLEMENTATION_H
+
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/base.h"
+#include "simdjson/implementation.h"
+#include "simdjson/internal/instruction_set.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+namespace simdjson {
+namespace lasx {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("lasx", "LoongArch ASX", internal::instruction_set::LASX) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_IMPLEMENTATION_H

include/simdjson/lasx/intrinsics.h

@@ -0,0 +1,14 @@
+#ifndef SIMDJSON_LASX_INTRINSICS_H
+#define SIMDJSON_LASX_INTRINSICS_H
+
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/lasx/base.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <lasxintrin.h>
+
+static_assert(sizeof(__m256i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for LoongArch ASX");
+
+#endif //  SIMDJSON_LASX_INTRINSICS_H

include/simdjson/lasx/numberparsing_defs.h

@@ -0,0 +1,41 @@
+#ifndef SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+#define SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+
+#ifndef SIMDJSON_CONDITIONAL_INCLUDE
+#include "simdjson/lasx/base.h"
+#include "simdjson/lasx/intrinsics.h"
+#include "simdjson/internal/numberparsing_tables.h"
+#endif // SIMDJSON_CONDITIONAL_INCLUDE
+
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace lasx
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_LASX_NUMBERPARSING_DEFS_H

include/simdjson/lasx/ondemand.h

@@ -0,0 +1,8 @@
+#ifndef SIMDJSON_LASX_ONDEMAND_H
+#define SIMDJSON_LASX_ONDEMAND_H
+
+#include "simdjson/lasx/begin.h"
+#include "simdjson/generic/ondemand/amalgamated.h"
+#include "simdjson/lasx/end.h"
+
+#endif // SIMDJSON_LASX_ONDEMAND_H