Do not store the lower bits mask in EliasFanoReader

Summary:
Computing the mask on access has negligible cost as it can be
hoisted out of the linear search loop, and furthermore on Haswell we
can use the the `BZHI` instruction.

I also experimented with `BEXTR` but it ended up being slower because
computing the pattern operand requires a shift and an or (it's
probably meant for when the pattern is precomputed).

Reviewed By: philippv

Differential Revision: D4976657

fbshipit-source-id: e4c4ca5f0a785595587e6d6ad4676f5b216291cf

diff --git a/folly/experimental/EliasFanoCoding.h b/folly/experimental/EliasFanoCoding.h
index f75fabde98c8ecd1e134cf627d9dfa5583d8d355..98ac516e22411c2ba5c8117faa92ea9745b05b40 100644 (file)
--- a/folly/experimental/EliasFanoCoding.h
+++ b/folly/experimental/EliasFanoCoding.h
@@ -28,6 +28,7 @@
 #include <limits>
 #include <type_traits>
 
+#include <folly/Assume.h>
 #include <folly/Bits.h>
 #include <folly/Likely.h>
 #include <folly/Portability.h>
@@ -526,7 +527,6 @@ class EliasFanoReader {
       : size_(list.size),
         lower_(list.lower),
         upper_(list),
-        lowerMask_((ValueType(1) << list.numLowerBits) - 1),
         numLowerBits_(list.numLowerBits) {
     DCHECK(Instructions::supported());
     // To avoid extra branching during skipTo() while reading
@@ -654,7 +654,10 @@ class EliasFanoReader {
     const size_t pos = i * numLowerBits_;
     const unsigned char* ptr = lower_ + (pos / 8);
     const uint64_t ptrv = folly::loadUnaligned<uint64_t>(ptr);
-    return lowerMask_ & (ptrv >> (pos % 8));
+    // This removes the branch in the fallback implementation of
+    // bzhi. The condition is verified at encoding time.
+    assume(numLowerBits_ < sizeof(ValueType) * 8);
+    return Instructions::bzhi(ptrv >> (pos % 8), numLowerBits_);
   }
 
   void iterateTo(ValueType value) {
@@ -671,7 +674,6 @@ class EliasFanoReader {
   size_t size_;
   const uint8_t* lower_;
   detail::UpperBitsReader<Encoder, Instructions> upper_;
-  const ValueType lowerMask_;
   ValueType value_ = kInvalidValue;
   ValueType lastValue_;
   uint8_t numLowerBits_;

diff --git a/folly/experimental/Instructions.h b/folly/experimental/Instructions.h
index a2754b8907c6acf498a7ae64ac2633bc44ca9e75..972b62d79a982bf2993a0472330f3adce8b494fd 100644 (file)
--- a/folly/experimental/Instructions.h
+++ b/folly/experimental/Instructions.h
@@ -72,6 +72,14 @@ struct Default {
     return (value >> start) &
         ((length == 64) ? (~0ULL) : ((1ULL << length) - 1ULL));
   }
+
+  // Clear high bits starting at position index.
+  static FOLLY_ALWAYS_INLINE uint64_t bzhi(uint64_t value, uint32_t index) {
+    if (index > 63) {
+      return 0;
+    }
+    return value & ((uint64_t(1) << index) - 1);
+  }
 };
 
 struct Nehalem : public Default {
@@ -94,12 +102,12 @@ struct Nehalem : public Default {
 
 struct Haswell : public Nehalem {
   static bool supported(const folly::CpuId& cpuId = {}) {
-    return Nehalem::supported(cpuId) && cpuId.bmi1();
+    return Nehalem::supported(cpuId) && cpuId.bmi1() && cpuId.bmi2();
   }
 
   static FOLLY_ALWAYS_INLINE uint64_t blsr(uint64_t value) {
 // BMI1 is supported starting with Intel Haswell, AMD Piledriver.
-// BLSR combines two instuctions into one and reduces register pressure.
+// BLSR combines two instructions into one and reduces register pressure.
 #if defined(__GNUC__) || defined(__clang__)
     // GCC and Clang won't inline the intrinsics.
     uint64_t result;
@@ -124,6 +132,18 @@ struct Haswell : public Nehalem {
     return result;
 #else
     return _bextr_u64(value, start, length);
+#endif
+  }
+
+  static FOLLY_ALWAYS_INLINE uint64_t bzhi(uint64_t value, uint32_t index) {
+#if defined(__GNUC__) || defined(__clang__)
+    // GCC and Clang won't inline the intrinsics.
+    const uint64_t index64 = index;
+    uint64_t result;
+    asm("bzhiq %2, %1, %0" : "=r"(result) : "r"(value), "r"(index64));
+    return result;
+#else
+    return _bzhi_u64(value, index);
 #endif
   }
 };