feat: add s390x architecture support (#1214)

* fix(lua): use native architecture when compiling lua for s390x.

Signed-off-by: iko1 <me@remotecpp.dev>

* feat(server): implement CompareFP for s390x architecture.

Signed-off-by: iko1 <me@remotecpp.dev>

* feat: implement validate_ascii_fast function variant for s390x arch.

Signed-off-by: iko1 <me@remotecpp.dev>

* fix: add comments before s390x vector operations

Signed-off-by: iko1 <me@remotecpp.dev>

* fix validate_ascii_fast function logic after CR comment

Signed-off-by: iko1 <me@remotecpp.dev>

* Revert "fix(lua): use native architecture when compiling lua for s390x."

This reverts commit 6cc5d8a.

* fix(lua): use native architecture when compiling lua for s390x.

Signed-off-by: iko1 <me@remotecpp.dev>

* refactor validate_ascii_fast function after CR comment

Signed-off-by: iko1 <me@remotecpp.dev>

* include vecintrin.h from sse_port.h rather the misleading filename

Signed-off-by: iko1 <me@remotecpp.dev>

---------

Signed-off-by: iko1 <me@remotecpp.dev>

patches/lua-v5.4.4.patch

@@ -12,10 +12,10 @@ index d42d14b7..75647e72 100644
  #define LUAI_MAXSTACK		15000
  #endif
 diff --git a/makefile b/makefile
-index d46e650c..e347e614 100644
+index d46e650c..c27e5677 100644
 --- a/makefile
 +++ b/makefile
-@@ -66,13 +66,23 @@ LOCAL = $(TESTS) $(CWARNS)
+@@ -66,13 +66,25 @@ LOCAL = $(TESTS) $(CWARNS)
 
 
  # enable Linux goodies
@@ -32,6 +32,8 @@ index d46e650c..e347e614 100644
 +OPTFLAGS= -march=sandybridge
 +else ifeq ($(uname_m), aarch64)
 +OPTFLAGS= -march=armv8.2-a+fp16+rcpc+dotprod+crypto
++else ifeq ($(uname_m), s390x)
++OPTFLAGS= -march=native
 +else
 +  $(error ERROR: unknown architecture $(uname_m))
 +endif

src/core/dash_internal.h

@@ -879,6 +879,34 @@ unsigned BucketBase<NUM_SLOTS, NUM_OVR>::UnsetStashPtr(uint8_t fp_hash, unsigned
   return res;
 }
 
+#ifdef __s390x__
+template <unsigned NUM_SLOTS, unsigned NUM_OVR>
+uint32_t BucketBase<NUM_SLOTS, NUM_OVR>::CompareFP(uint8_t fp) const {
+  static_assert(FpArray{}.size() <= 16);
+  vector unsigned char v1;
+
+  // Replicate 16 times fp to key_data.
+  for (int i = 0; i < 16; i++) {
+    v1[i] = fp;
+  }
+
+  // Loads 16 bytes of src into seg_data.
+  vector unsigned char v2 = vec_load_len(finger_arr_.data(), 16);
+
+  // compare 1-byte vectors seg_data and key_data, dst[i] := ( a[i] == b[i] ) ? 0xFF : 0.
+  vector bool char rv_mask = vec_cmpeq(v1, v2);
+
+  // collapses 16 msb bits from each byte in rv_mask into mask.
+  int mask = 0;
+  for (int i = 0; i < 16; i++) {
+    if (rv_mask[i]) {
+      mask |= 1 << i;
+    }
+  }
+
+  return mask;
+}
+#else
 template <unsigned NUM_SLOTS, unsigned NUM_OVR>
 uint32_t BucketBase<NUM_SLOTS, NUM_OVR>::CompareFP(uint8_t fp) const {
   static_assert(FpArray{}.size() <= 16);
@@ -898,6 +926,7 @@ uint32_t BucketBase<NUM_SLOTS, NUM_OVR>::CompareFP(uint8_t fp) const {
   // Note: Last 2 operations can be combined in skylake with _mm_cmpeq_epi8_mask.
   return mask;
 }
+#endif
 
 // Bucket slot array goes from left to right: [x, x, ...]
 // Shift right vacates the first slot on the left by shifting all the elements right and

src/core/detail/bitpacking.cc

@@ -104,6 +104,42 @@ static inline pair<const char*, uint8_t*> simd_variant2_pack(const char* ascii,
 // See https://github.com/lemire/fastvalidate-utf-8/
 // The function returns true (1) if all chars passed in src are
 // 7-bit values (0x00..0x7F). Otherwise, it returns false (0).
+#ifdef __s390x__
+bool validate_ascii_fast(const char* src, size_t len) {
+  size_t i = 0;
+
+  // Initialize a vector in which all the elements are set to zero.
+  vector unsigned char has_error = vec_splat_s8(0);
+  if (len >= 16) {
+    for (; i <= len - 16; i += 16) {
+      // Load 16 bytes from buffer into a vector.
+      vector unsigned char current_bytes = vec_load_len((signed char*)(src + i), 16);
+      // Perform a bitwise OR operation between the current and the previously loaded contents.
+      has_error = vec_orc(has_error, current_bytes);
+    }
+  }
+
+  // Initialize a vector in which all the elements are set to an invalid ASCII value.
+  vector unsigned char rep_invalid_values = vec_splat_s8(0x80);
+
+  // Perform bitwise AND-complement operation between two vectors.
+  vector unsigned char andc_result = vec_andc(rep_invalid_values, has_error);
+
+  // Tests whether any of corresponding elements of the given vectors are not equal.
+  // After the bitwise operation, both vectors should be equal if ASCII values.
+  if (!vec_all_eq(rep_invalid_values, andc_result)) {
+    return false;
+  }
+
+  for (; i < len; i++) {
+    if (src[i] & 0x80) {
+      return false;
+    }
+  }
+
+  return true;
+}
+#else
 bool validate_ascii_fast(const char* src, size_t len) {
   size_t i = 0;
   __m128i has_error = _mm_setzero_si128();
@@ -123,6 +159,7 @@ bool validate_ascii_fast(const char* src, size_t len) {
 
   return !error_mask;
 }
+#endif
 
 // len must be at least 16
 void ascii_pack(const char* ascii, size_t len, uint8_t* bin) {

src/core/sse_port.h

@@ -5,13 +5,16 @@
 #pragma once
 #if defined(__aarch64__)
 #include "base/sse2neon.h"
+#elif defined(__s390x__)
+#include <vecintrin.h>
 #else
 #include <emmintrin.h>
 #include <tmmintrin.h>
 #endif
 
 namespace dfly {
 
+#ifndef __s390x__
 inline __m128i mm_loadu_si128(const __m128i* ptr) {
 #if defined(__aarch64__)
   __m128i res;
@@ -22,5 +25,6 @@ inline __m128i mm_loadu_si128(const __m128i* ptr) {
   return _mm_loadu_si128(ptr);
 #endif
 }
+#endif
 
 }  // namespace dfly