Use less compiler architecture builtins and more vector code.

1 files changed, 30 insertions, 91 deletions

diff --git a/streamingvisitors/src/vespa/vsm/searcher/fold.cpp b/streamingvisitors/src/vespa/vsm/searcher/fold.cpp
index bd2392d3ad6..4542c98c25c 100644
--- a/streamingvisitors/src/vespa/vsm/searcher/fold.cpp
+++ b/streamingvisitors/src/vespa/vsm/searcher/fold.cpp
@@ -2,33 +2,32 @@
 //
 #include "fold.h"
 
+using search::v16qi;
+using search::v2di;
+
 namespace vsm {
+namespace {
 
+constexpr v16qi G_0    = { '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1 };
+constexpr v16qi G_9    = { '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9' };
+constexpr v16qi G_a    = { 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1 };
+constexpr v16qi G_z    = { 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z' };
+constexpr v16qi G_8bit = { (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4,
+                            (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4 };
+constexpr v2di  G_lowCase = { 0x2020202020202020ULL, 0x2020202020202020ULL };
+}
 const unsigned char * sse2_foldaa(const unsigned char * toFoldOrg, size_t sz, unsigned char * foldedOrg)
 {
-  typedef char v16qi __attribute__ ((__vector_size__(16)));
-  typedef long long v2di  __attribute__ ((__vector_size__(16)));
-  static v16qi _G_0    = { '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1 };
-  static v16qi _G_9    = { '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9' };
-  static v16qi _G_a    = { 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1 };
-  static v16qi _G_z    = { 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z' };
-  static v16qi _G_8bit = { (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4,
-                           (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4 };
-  static v2di  _G_lowCase = { 0x2020202020202020ULL, 0x2020202020202020ULL };
-  const v16qi *toFold = reinterpret_cast<const v16qi *>(toFoldOrg);
+  const auto * toFold = reinterpret_cast<const v16qi *>(toFoldOrg);
   v2di * folded =  reinterpret_cast<v2di *>(foldedOrg);
   size_t i=0;
   for (size_t m=sz/16; i < m; i++)
   {
-#ifndef __INTEL_COMPILER
     int nonAscii = __builtin_ia32_pmovmskb128(toFold[i]);
     if (nonAscii)
     {
-#ifdef __clang__
-      v16qi non8Mask = _G_8bit > toFold[i];
-#else
-      v16qi non8Mask = __builtin_ia32_pcmpgtb128(_G_8bit, toFold[i]);
-#endif
+      v16qi non8Mask = G_8bit > toFold[i];
+
       int non8bit = __builtin_ia32_pmovmskb128(non8Mask);
       if (non8bit)
       {
@@ -36,74 +35,35 @@ const unsigned char * sse2_foldaa(const unsigned char * toFoldOrg, size_t sz, un
       }
       break;
     }
-#ifdef __clang__
-    v16qi _0     = toFold[i] > _G_0;
-    v16qi _z     = toFold[i] > _G_z;
+    v16qi _0     = toFold[i] > G_0;
+    v16qi _z     = toFold[i] > G_z;
     v2di  _0_z   = v2di(_0) ^ v2di(_z);
     v2di  toLow  = _0_z & v2di(toFold[i]);
-    v16qi low    = v16qi(toLow | _G_lowCase);
-    _0           = low > _G_0;
-    v16qi _9     = low > _G_9;
-    v16qi _a     = low > _G_a;
-    _z           = low > _G_z;
+    auto  low    = v16qi(toLow | G_lowCase);
+    _0           = low > G_0;
+    v16qi _9     = low > G_9;
+    v16qi _a     = low > G_a;
+    _z           = low > G_z;
     v2di  _0_9_m = v2di(_0) ^ v2di(_9);
     v2di  _a_z_m = v2di(_a) ^ v2di(_z);
     v2di  _0_9   = _0_9_m & v2di(low);
     v2di  _a_z   = _a_z_m & v2di(low);
     folded[i]    = _0_9 | _a_z;
-#else
-    v16qi _0     = __builtin_ia32_pcmpgtb128(toFold[i], _G_0);
-    v16qi _z     = __builtin_ia32_pcmpgtb128(toFold[i], _G_z);
-    v2di  _0_z   = __builtin_ia32_pxor128(v2di(_0), v2di(_z));
-    v2di  toLow  = __builtin_ia32_pand128(_0_z, v2di(toFold[i]));
-    v16qi low    = v16qi(__builtin_ia32_por128(toLow, _G_lowCase));
-    _0           = __builtin_ia32_pcmpgtb128(low, _G_0);
-    v16qi _9     = __builtin_ia32_pcmpgtb128(low, _G_9);
-    v16qi _a     = __builtin_ia32_pcmpgtb128(low, _G_a);
-    _z           = __builtin_ia32_pcmpgtb128(low, _G_z);
-    v2di  _0_9_m = __builtin_ia32_pxor128(v2di(_0), v2di(_9));
-    v2di  _a_z_m = __builtin_ia32_pxor128(v2di(_a), v2di(_z));
-    v2di  _0_9   = __builtin_ia32_pand128(_0_9_m, v2di(low));
-    v2di  _a_z   = __builtin_ia32_pand128(_a_z_m, v2di(low));
-    folded[i]    = __builtin_ia32_por128(_0_9, _a_z);
-#endif
-#else
-#   warning "Intel's icc compiler does not like __builtin_ia32_pxor128"
-    LOG_ABORT("should not be reached");
-#endif
   }
   return toFoldOrg+i*16;
 }
 
 const unsigned char * sse2_foldua(const unsigned char * toFoldOrg, size_t sz, unsigned char * foldedOrg)
 {
-  typedef char v16qi __attribute__ ((__vector_size__(16)));
-  typedef long long v2di  __attribute__ ((__vector_size__(16)));
-  static v16qi _G_0    = { '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1, '0'-1 };
-  static v16qi _G_9    = { '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9', '9' };
-  static v16qi _G_a    = { 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1, 'a'-1 };
-  static v16qi _G_z    = { 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z', 'z' };
-  static v16qi _G_8bit = { (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4,
-                           (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4, (char)0xc4 };
-  static v2di  _G_lowCase = { 0x2020202020202020ULL, 0x2020202020202020ULL };
   v2di * folded =  reinterpret_cast<v2di *>(foldedOrg);
   size_t i=0;
   for (size_t m=sz/16; i < m; i++)
   {
-#ifndef __INTEL_COMPILER
-#ifdef __clang__
     v16qi current = __builtin_ia32_lddqu(reinterpret_cast<const char *>(&toFoldOrg[i*16]));
-#else
-    v16qi current = __builtin_ia32_loaddqu(reinterpret_cast<const char *>(&toFoldOrg[i*16]));
-#endif
     int nonAscii = __builtin_ia32_pmovmskb128(current);
     if (nonAscii)
     {
-#ifdef __clang__
-      v16qi non8Mask = _G_8bit > current;
-#else
-      v16qi non8Mask = __builtin_ia32_pcmpgtb128(_G_8bit, current);
-#endif
+      v16qi non8Mask = G_8bit > current;
       int non8bit = __builtin_ia32_pmovmskb128(non8Mask);
       if (non8bit)
       {
@@ -111,41 +71,20 @@ const unsigned char * sse2_foldua(const unsigned char * toFoldOrg, size_t sz, un
       }
       break;
     }
-#ifdef __clang__
-    v16qi _0     = current > _G_0;
-    v16qi _z     = current > _G_z;
+    v16qi _0     = current > G_0;
+    v16qi _z     = current > G_z;
     v2di  _0_z   = v2di(_0) ^ v2di(_z);
     v2di  toLow  = _0_z & v2di(current);
-    v16qi low    = v16qi(toLow | _G_lowCase);
-    _0           = low > _G_0;
-    v16qi _9     = low > _G_9;
-    v16qi _a     = low > _G_a;
-    _z           = low > _G_z;
+    auto  low    = v16qi(toLow | G_lowCase);
+    _0           = low > G_0;
+    v16qi _9     = low > G_9;
+    v16qi _a     = low > G_a;
+    _z           = low > G_z;
     v2di  _0_9_m = v2di(_0) ^ v2di(_9);
     v2di  _a_z_m = v2di(_a) ^ v2di(_z);
     v2di  _0_9   = _0_9_m & v2di(low);
     v2di  _a_z   = _a_z_m & v2di(low);
     folded[i]    = _0_9 | _a_z;
-#else
-    v16qi _0     = __builtin_ia32_pcmpgtb128(current, _G_0);
-    v16qi _z     = __builtin_ia32_pcmpgtb128(current, _G_z);
-    v2di  _0_z   = __builtin_ia32_pxor128(v2di(_0), v2di(_z));
-    v2di  toLow  = __builtin_ia32_pand128(_0_z, v2di(current));
-    v16qi low    = v16qi(__builtin_ia32_por128(toLow, _G_lowCase));
-    _0           = __builtin_ia32_pcmpgtb128(low, _G_0);
-    v16qi _9     = __builtin_ia32_pcmpgtb128(low, _G_9);
-    v16qi _a     = __builtin_ia32_pcmpgtb128(low, _G_a);
-    _z           = __builtin_ia32_pcmpgtb128(low, _G_z);
-    v2di  _0_9_m = __builtin_ia32_pxor128(v2di(_0), v2di(_9));
-    v2di  _a_z_m = __builtin_ia32_pxor128(v2di(_a), v2di(_z));
-    v2di  _0_9   = __builtin_ia32_pand128(_0_9_m, v2di(low));
-    v2di  _a_z   = __builtin_ia32_pand128(_a_z_m, v2di(low));
-    folded[i]    = __builtin_ia32_por128(_0_9, _a_z);
-#endif
-#else
-#   warning "Intel's icc compiler does not like __builtin_ia32_pxor128"
-    LOG_ABORT("should not be reached");
-#endif
   }
   return toFoldOrg+i*16;
 }