Commit 7213df65 authored by Christopher Dykes's avatar Christopher Dykes Committed by Facebook Github Bot

Always use inline-asm version on GCC/Clang

Summary:
Because the intrinsic version requires explicitly adding `__target__` attributes, which results in things not being inlined. Although the code generated with the `__target__` attribute is strictly better, ensuring it's applied on all the relevant functions is error-prone, so just use the inline assembly version for GCC/Clang so that it can be inlined elsewhere. MSVC will inline the intrinsic version without any issue.
This also marks the functions as `ALWAYS_INLINE`, as the diff that is getting reverted made that change as well.

Reviewed By: yfeldblum, philippv, ot

Differential Revision: D3963935

fbshipit-source-id: 47175d64e7be351eb455a4d053b91ce9392bf152
parent f588dc4a
......@@ -18,17 +18,14 @@
#include <glog/logging.h>
#ifdef _MSC_VER
#include <immintrin.h>
#endif
#include <folly/CpuId.h>
#include <folly/Portability.h>
#include <folly/portability/Builtins.h>
#if defined(__GNUC__) || defined(__clang__)
// For compilers supporting AT&T assembly syntax.
#define FOLLY_INSTRUCTIONS_SUPPORTED 1
#else
#define FOLLY_INSTRUCTIONS_SUPPORTED 0
#endif
namespace folly { namespace compression { namespace instructions {
// NOTE: It's recommended to compile EF coding with -msse4.2, starting
......@@ -59,8 +56,6 @@ struct Default {
}
};
#if FOLLY_INSTRUCTIONS_SUPPORTED
struct Nehalem : public Default {
static bool supported(const folly::CpuId& cpuId = {}) {
return cpuId.popcnt();
......@@ -68,9 +63,14 @@ struct Nehalem : public Default {
static FOLLY_ALWAYS_INLINE uint64_t popcount(uint64_t value) {
// POPCNT is supported starting with Intel Nehalem, AMD K10.
#if defined(__GNUC__) || defined(__clang__)
// GCC and Clang won't inline the intrinsics.
uint64_t result;
asm ("popcntq %1, %0" : "=r" (result) : "r" (value));
return result;
#else
return _mm_popcnt_u64(value);
#endif
}
};
......@@ -82,17 +82,15 @@ struct Haswell : public Nehalem {
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.
#if defined(__GNUC__) || defined(__clang__)
// GCC and Clang won't inline the intrinsics.
uint64_t result;
asm ("blsrq %1, %0" : "=r" (result) : "r" (value));
return result;
#else
return _blsr_u64(value);
#endif
}
};
#else // FOLLY_INSTRUCTIONS_SUPPORTED
struct Nehalem : public Default {};
struct Haswell : public Nehalem {};
#endif // FOLLY_INSTRUCTIONS_SUPPORTED
}}} // namespaces
......@@ -63,11 +63,11 @@ inline uint64_t select64(uint64_t x, uint64_t k) {
return place + detail::kSelectInByte[((x >> place) & 0xFF) | (byteRank << 8)];
}
#if FOLLY_INSTRUCTIONS_SUPPORTED
template <>
FOLLY_ALWAYS_INLINE uint64_t
select64<compression::instructions::Haswell>(uint64_t x, uint64_t k) {
#if defined(__GNUC__) || defined(__clang__)
// GCC and Clang won't inline the intrinsics.
uint64_t result = uint64_t(1) << k;
asm("pdep %1, %0, %0\n\t"
......@@ -76,8 +76,9 @@ select64<compression::instructions::Haswell>(uint64_t x, uint64_t k) {
: "r"(x));
return result;
#else
return _tzcnt_u64(_pdep_u64(1ULL << k, x));
#endif
}
#endif // FOLLY_INSTRUCTIONS_SUPPORTED
} // namespace folly
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment