[SSE2] poor performance with emulated _mm_extract_epi8

[llvmbot]

Bugzilla Link	27737
Version	3.8
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@d0k,@RKSimon,@rotateright

Extended Description

I have some code that heavy uses _mm_extract_epi8().
Then I build with clang 3.8 and -msse2 (without -msse4.1) then program work very slow.

To build with GCC and clang 3.4, 3.6, 3.7 I use macro:

#ifndef _mm_extract_epi8 /* SSE4.1 required. */
#define _mm_extract_epi8(__xmm, __n)					\
    ((_mm_extract_epi16(__xmm, ((__n) >> 1)) >> (8 * ((__n) & 1))) & 0xff)
#endif

Test results:

AMD Athlon(tm) 5350 APU with Radeon(tm) R3      (2050.04-MHz K8-class CPU)
GCC:		20391006000 (SSE4.1) /  20116413000 (SSE2)
clang 3.8:	22329895000 (SSE4.1) / 117304135000 (SSE2) !!!
clang 3.7:	22367008000 (SSE4.1) /  25542571000 (SSE2)
clang 3.6:	22306648000 (SSE4.1) /  25914115000 (SSE2)
clang 3.4:	23684031000 (SSE4.1) /  25914115000 (SSE2)

Intel(R) Core(TM)2 Duo CPU     E8400  @ 3.00GHz (2999.72-MHz K8-class CPU)
GCC:		12031595000 (SSE4.1) / 12011303000 (SSE2)
clang 3.8:	12431116000 (SSE4.1) / 73035466000 (SSE2) !!!
clang 3.7:	12458839000 (SSE4.1) / 13317058000 (SSE2)
clang 3.6:	12462181000 (SSE4.1) / 14119683000 (SSE2)
clang 3.4:	13555167000 (SSE4.1) / 13178893000 (SSE2)

[d0k]

Without additional info about the code that's actually being run this bug report isn't useful.

[RKSimon]

Looking at this now.

The problem is that we're currently missing pre-SSE41 custom lowering for extractelement i8 - so we end up spilling the vector to stack and loading the byte (which can be particularly costly).

[llvmbot]

Without additional info about the code that's actually being run this bug
report isn't useful.

http://www.netlab.linkpc.net/download/software/SDK/core/include/gost3411-2012.h

	struct rusage ru_st, ru_end;
	size_t i, cnt = 10;
	gost3411_2012_ctx_t ctx;
	uint8_t digest[GOST3411_2012_HASH_MAX_SIZE];
	uint8_t *data;
	size_t data_size = (100 * 1024 * 1024);

	data = malloc((data_size + 32));
	memset(data, 0xaa, (data_size + 32));
	data += 1;

	getrusage(RUSAGE_SELF, &ru_st);
	for (i = 0; i < cnt; i ++) {
		gost3411_2012_init(512, &ctx);
		gost3411_2012_update(&ctx, data, data_size);
		gost3411_2012_final(&ctx, digest);
	}
	getrusage(RUSAGE_SELF, &ru_end);

But I made fix in my code:

#ifndef __SSE4_1__ /* SSE4.1 required. */
#undef _mm_extract_epi8 /* CLang 3.8 perfomance fix. */
#define _mm_extract_epi8(__xmm, __n)					\
    (0xff & (_mm_extract_epi16(__xmm, ((__n) >> 1)) >> (8 * ((__n) & 0x01))))
#endif

comment/delete it to reproduce bug.

[RKSimon]

D29841/rL297568 improved v16i8 extraction if its the only use of the vector - multiple extractions still go via a stack spill.

[RKSimon]

resolving old issue - this was further improved by D156350 / 071671e - which would extract multiple elements if they could came from a common wider extraction using MOVD/PEXTRW - otherwise stack spill would still be the better approach