Use intrinsics rather than inline assembly where possible

Summary:
I would switch these to just use the intrinsic functions, but GCC 4.8 doesn't support them.
MSVC supports the intrinsics, which is the primary reason for the switch.

Reviewed By: yfeldblum

Differential Revision: D3265572

fb-gh-sync-id: f12d16b8f64cf48adcb97958d7ae88d56689a96f
fbshipit-source-id: f12d16b8f64cf48adcb97958d7ae88d56689a96f

diff --git a/folly/Portability.h b/folly/Portability.h
index c9616ea7caa23dcaa304007be149877249218b88..cbf1b71c3ef2ed2713307caf6c2b1915fad64944 100644 (file)
--- a/folly/Portability.h
+++ b/folly/Portability.h
@@ -92,6 +92,13 @@ constexpr bool kHasUnalignedAccess = false;
 # define FOLLY_ALWAYS_INLINE inline
 #endif
 
+// target
+#ifdef _MSC_VER
+# define FOLLY_TARGET_ATTRIBUTE(target)
+#else
+# define FOLLY_TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
+#endif
+
 // detection for 64 bit
 #if defined(__x86_64__) || defined(_M_X64)
 # define FOLLY_X64 1

diff --git a/folly/experimental/Instructions.h b/folly/experimental/Instructions.h
index 97386ff16b5311cbddd7c9b48bd9cee95b45702c..6a1d6ed653be3819c6664e6c4f6fbb3cc3310475 100644 (file)
--- a/folly/experimental/Instructions.h
+++ b/folly/experimental/Instructions.h
@@ -17,8 +17,10 @@
 #pragma once
 
 #include <glog/logging.h>
+#include <immintrin.h>
 
 #include <folly/CpuId.h>
+#include <folly/portability/Builtins.h>
 
 namespace folly { namespace compression { namespace instructions {
 
@@ -52,11 +54,17 @@ struct Nehalem : public Default {
   static bool supported(const folly::CpuId& cpuId = {}) {
     return cpuId.popcnt();
   }
-  static inline uint64_t popcount(uint64_t value) {
+  static inline uint64_t popcount(uint64_t value)
+      FOLLY_TARGET_ATTRIBUTE("popcnt") {
     // POPCNT is supported starting with Intel Nehalem, AMD K10.
+#if defined(__GNUC__) && !__GNUC_PREREQ(4, 9)
+    // GCC 4.8 doesn't support the intrinsics.
     uint64_t result;
     asm ("popcntq %1, %0" : "=r" (result) : "r" (value));
     return result;
+#else
+    return _mm_popcnt_u64(value);
+#endif
   }
 };
 
@@ -64,12 +72,17 @@ struct Haswell : public Nehalem {
   static bool supported(const folly::CpuId& cpuId = {}) {
     return Nehalem::supported(cpuId) && cpuId.bmi1();
   }
-  static inline uint64_t blsr(uint64_t value) {
+  static inline uint64_t blsr(uint64_t value) FOLLY_TARGET_ATTRIBUTE("bmi") {
     // BMI1 is supported starting with Intel Haswell, AMD Piledriver.
     // BLSR combines two instuctions into one and reduces register pressure.
+#if defined(__GNUC__) && !__GNUC_PREREQ(4, 9)
+    // GCC 4.8 doesn't support the intrinsics.
     uint64_t result;
     asm ("blsrq %1, %0" : "=r" (result) : "r" (value));
     return result;
+#else
+    return _blsr_u64(value);
+#endif
   }
 };
 

diff --git a/folly/experimental/Select64.h b/folly/experimental/Select64.h
index 9c2e03f90aebc788e9cb992f36825092d8574d63..5d0fcaa5b51e69cec887dc4d6735cfc666f78d0f 100644 (file)
--- a/folly/experimental/Select64.h
+++ b/folly/experimental/Select64.h
@@ -62,9 +62,15 @@ inline uint64_t select64(uint64_t x, uint64_t k) {
   return place + detail::kSelectInByte[((x >> place) & 0xFF) | (byteRank << 8)];
 }
 
+template <>
+uint64_t select64<compression::instructions::Haswell>(uint64_t x, uint64_t k)
+  FOLLY_TARGET_ATTRIBUTE("bmi,bmi2");
+
 template <>
 inline uint64_t select64<compression::instructions::Haswell>(uint64_t x,
                                                              uint64_t k) {
+#if defined(__GNUC__) && !__GNUC_PREREQ(4, 9)
+  // GCC 4.8 doesn't support the intrinsics.
   uint64_t result = uint64_t(1) << k;
 
   asm("pdep %1, %0, %0\n\t"
@@ -73,6 +79,9 @@ inline uint64_t select64<compression::instructions::Haswell>(uint64_t x,
       : "r"(x));
 
   return result;
+#else
+  return _tzcnt_u64(_pdep_u64(x, 1ULL << k));
+#endif
 }
 
 } // namespace folly