ilyakurdyukov/e2k-ports

e2k-ports

Performance patches and build fixes for Elbrus (e2k) architecture.

This is my personal repository so that patches won't get lost.

Elbrus porting cheat sheet:

Elbrus 2000 (aka e2k) is a 64-bit little-endian architecture.
The compiler is mostly GCC compatible (defines __GNUC__), EDG frontend.

detection

shell: uname -m returns e2k
cmake: if({CMAKE_SYSTEM_PROCESSOR} STREQUAL "e2k")
C preprocessor: if defined(__e2k__)
compiler version: if __LCC__ = 125 and __LCC_MINOR__ = 9 then it's "LCC 1.25.09"
architecture version: defined in __iset__ (less than 3 is obsolete, 6 is the latest at the moment)

intrinsics

MMX, SSE2, SSSE3, SSE4.1* - native support
AVX, AVX2 - supported, but not recommended, uses too much CPU registers
SSE4.2 and _mm_dp_ps (from SSE4.1) - emulated, slow, do not use

The compiler enables MMX to AVX2 support by default, pass -mno-avx (-mno-sse4.2) if code depends on the presence of macros (e.g. #if defined(__AVX2__)).

builtins

__sync*, __atomic* - supported by the compiler
count leading/trailing zeros - supported (__builtin_clz, __builtin_ctz)
memory fence - supported (need to include x86intrin.h first)
- __builtin_ia32_mfence, __builtin_ia32_lfence, __builtin_ia32_sfence

cpuid

Use compile time CPU detection, select the best SIMD up to SSE4.1.

rdtsc

#include <x86intrin.h>
uint64_t time = __rdtsc();
// same: unsigned aux; uint64_t time = __rdtscp(&aux);

useful pragmas

_Pragma("name") - to use from macros.

Use before the loop:

#pragma ivdep - ignore data dependencies inside the loop
#pragma unroll(n) - unroll cycle N times

restrict

Using the restrict keyword is good for performance, but note that it is ignored by the LCC if you're using vector load/store intrinsics such as _mm_load_si128(). For code with vector intrinsics use #pragma ivdep.

makecontext

Instead of makecontext(ctx, ...) use makecontext_e2k(ctx, ...), returns a negative integer on error. Allocates extra resources that need to be freed using freecontext_e2k(ctx).

nop

Use __asm__ __volatile__ ("nop") or _mm_pause() for a little delay.

clearing the instruction cache

The GNUC standard function __clear_cache(char *begin, char *end) works correctly since LCC 1.25.18, LCC 1.26.04. This function is available in previous versions, but does nothing.

inline

If it's crucial to performance, then use __attribute__((__always_inline__)) inline rather than just inline. Because when using large or complicated inline functions, the LCC compiler may decide not to inline them.

avoid if possible

The GNUC C extension Labels as Values is available in the LCC, but performance is worse than using a simple switch/case.

The GNUC Vector Extension is also available in LCC, but poorly implemented and its performance is very bad.

ilyakurdyukov / e2k-ports