From: Christopher Dykes <cdykes@fb.com>
Date: Tue, 24 May 2016 20:57:25 +0000 (-0700)
Subject: Update generate_varint_tables to support MSVC
X-Git-Tag: 2016.07.26~206
X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=42c850c04cc6a72087860ebd2651ab1b04a3f001;p=folly.git

Update generate_varint_tables to support MSVC

Summary: MSVC is a massive pain in this respect, and, after testing many workarounds, and only ending up with it just emitting a dynamic initializer for them, it's easier to just generate the table as a `uint64_t` and load it explicitly.

Reviewed By: yfeldblum

Differential Revision: D3270226

fbshipit-source-id: 77bc84e58d393373de05a28a30dcb80b66c09c9f
---

diff --git a/folly/GroupVarint.h b/folly/GroupVarint.h
index b17d15f7..2dfb09b6 100644
--- a/folly/GroupVarint.h
+++ b/folly/GroupVarint.h
@@ -37,7 +37,7 @@
 #include <nmmintrin.h>
 namespace folly {
 namespace detail {
-extern const __m128i groupVarintSSEMasks[];
+alignas(16) extern const uint64_t groupVarintSSEMasks[];
 }  // namespace detail
 }  // namespace folly
 #endif
@@ -196,7 +196,8 @@ class GroupVarint<uint32_t> : public detail::GroupVarintBase<uint32_t> {
   static const char* decode(const char* p, uint32_t* dest) {
     uint8_t key = p[0];
     __m128i val = _mm_loadu_si128((const __m128i*)(p+1));
-    __m128i mask = detail::groupVarintSSEMasks[key];
+    __m128i mask =
+        _mm_load_si128((const __m128i*)&detail::groupVarintSSEMasks[key * 2]);
     __m128i r = _mm_shuffle_epi8(val, mask);
     _mm_storeu_si128((__m128i*)dest, r);
     return p + detail::groupVarintLengths[key];
@@ -210,7 +211,8 @@ class GroupVarint<uint32_t> : public detail::GroupVarintBase<uint32_t> {
                             uint32_t* c, uint32_t* d) {
     uint8_t key = p[0];
     __m128i val = _mm_loadu_si128((const __m128i*)(p+1));
-    __m128i mask = detail::groupVarintSSEMasks[key];
+    __m128i mask =
+        _mm_load_si128((const __m128i*)&detail::groupVarintSSEMasks[key * 2]);
     __m128i r = _mm_shuffle_epi8(val, mask);
 
     // Extracting 32 bits at a time out of an XMM register is a SSE4 feature
diff --git a/folly/build/generate_varint_tables.py b/folly/build/generate_varint_tables.py
index 96bfd1d1..bf082d99 100755
--- a/folly/build/generate_varint_tables.py
+++ b/folly/build/generate_varint_tables.py
@@ -56,15 +56,11 @@ def generate(f):
 
 #include <stdint.h>
 
-#if (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2)
-#include <x86intrin.h>
-#endif
-
 namespace folly {
 namespace detail {
 
 #if (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2)
-extern const __m128i groupVarintSSEMasks[] = {
+alignas(16) extern const uint64_t groupVarintSSEMasks[512] = {
 """)
 
     # Compute SSE masks
@@ -81,8 +77,8 @@ extern const __m128i groupVarintSSEMasks[] = {
             # 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"
         "#endif /*#if (FOLLY_X64 || defined(__i386__)) && (FOLLY_SSE >= 2)*/\n"