Compiler versions

I used compilers installed on amazing godbolt.org. The compilers versions:

• g++ (GCC-Explorer-Build) 9.0.0 20190120 (experimental)
• clang version 9.0.0 (trunk 352942)
• icc (ICC) 19.0.1.144 20181018
• Microsoft (R) C/C++ Optimizing Compiler Version 19.16.27023.1 for x86

Compiler flags

GCC, Clang:

• AVX2: -O3 -mavx2
• AVX512: -O3 -mavx512f -mavx512dq -mavx512bw -mavx512vbmi -mavx512vbmi2 -mavx512vl

ICC:

• AVX2: -O3 -axAVX2 -xAVX2
• AVX512: -O3 -mavx512f -mavx512dq -mavx512bw -mavx512vbmi -mavx512vbmi2 -mavx512vl

MSVC:

• AVX2: /O2 /arch:AVX2
• AVX512: /O2 /arch:AVX512

Results

algorithm	procedure	GCC		Clang		ICC		MSVC
		AVX2	AVX512	AVX2	AVX512	AVX2	AVX512	AVX2	AVX512
accumulate — custom	accumulate_custom_epi8	no	no	no	no	no	no	no	no[1]
	accumulate_custom_epi32	yes	yes	yes	yes	yes	yes	yes	no[1]
accumulate — default	accumulate_epi8	yes	yes	yes	yes	no	yes	yes	no[1]
	accumulate_epi32	yes	yes	yes	yes	yes	yes	yes	no[1]
all_of	all_of_epi8	no	no	no	no	no	no	no	no
	all_of_epi32	no	no	no	no	no	no	no	no
any_of	any_of_epi8	no	no	no	no	no	no	no	no
	any_of_epi32	no	no	no	no	no	no	no	no
copy	copy_epi8	no	no	no	no	no	no	no	no
	copy_epi32	no	no	no	no	no	no	no	no
copy_if	copy_if_epi8	no	no	no	no	no	no	no	no
	copy_if_epi32	no	no	no	no	no	no	no	no
count	count_epi8	yes	yes	yes	yes	yes	yes	yes	no[1]
	count_epi32	yes	yes	yes	yes	yes	yes	yes	no[1]
count_if	count_if_epi8	yes	yes	yes	yes	yes	yes	no	no
	count_if_epi32	yes	yes	yes	yes	yes	yes	no	no
fill	fill_epi8	no	no	no	no	yes	yes	no	no
	fill_epi32	yes	yes	yes	yes	yes	yes	no	no
find	find_epi8	no	no	no	no	no	no	no[2]	no[2]
	find_epi32	no	no	no	no	no	no	no	no
find_if	find_if_epi8	no	no	no	no	no	no	no	no
	find_if_epi32	no	no	no	no	no	no	no	no
is_sorted	is_sorted_epi8	no	no	no	no	no	no	no	no
	is_sorted_epi32	no	no	no	no	no	no	no	no
none_of	none_of_epi8	no	no	no	no	no	no	no	no
	none_of_epi32	no	no	no	no	no	no	no	no
remove	remove_epi8	no	no	no	no	no	no	no	no
	remove_epi32	no	no	no	no	no	no	no	no
remove_if	remove_if_epi8	no	no	no	no	no	no	no	no
	remove_if_epi32	no	no	no	no	no	no	no	no
replace	replace_epi8	no	yes	no[3]	yes	yes	yes	no	no
	replace_epi32	yes	yes	yes	yes	yes	yes	no	no
replace_if	replace_if_epi8	no	yes	no	yes	yes	yes	no	no
	replace_if_epi32	yes	no	no	no	yes	yes	no	no
reverse	reverse_epi8	yes	yes	no	no	no	no	no[4]	no[4]
	reverse_epi32	yes	yes	no	no	no	no	no[5]	no[5]
transform — abs	transform_abs_epi8	yes	yes	yes	yes	yes	yes	yes	no[1]
	transform_abs_epi32	yes	yes	yes	yes	yes	yes	yes	no[1]
transform — increment	transform_inc_epi8	yes	yes	yes	yes	yes	yes	yes	no[1]
	transform_inc_epi32	yes	yes	yes	yes	yes	yes	yes	no[1]
transform — negation	transform_neg_epi8	yes	yes	yes	yes	yes	yes	no	no
	transform_neg_epi32	yes	yes	yes	yes	yes	yes	no	no
unique	unique_epi8	no	no	no	no	no	no	no	no
	unique_epi32	no	no	no	no	no	no	no	no

[1]	(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) emits AVX2 code

[2]	(1, 2) calls memchr

[3]	loads input's chunk into a vector register, but all comparisons and stores are scalar

[4]	(1, 2) calls ___std_reverse_trivially_swappable_1

[5]	(1, 2) calls ___std_reverse_trivially_swappable_4

---

A few notes:

• In most compilers std::fill for int8_t ends up with memset invocation.

• MSVC does not emit AVX512 instruction, only AVX2.

• I was wondering why std::replace for int8_t cannot be vectorized for AVX2 targets. Following code would do the job:

  m = _mm256_cmpeq_epi8(x, _mm256_set1_epi8('/'));
  x = _mm256_blendv_epi8(x, _mm256_set1_epi8('_'), m);
  

  Travis Downs pointed out that such code behaves differently than original replace, as it always overwrites all bytes, regardless comparison results. To make things more interesting, ICC vectorizes std::replace using the above approach, which seems to be a compiler bug.

  However, if we rewrite replace algorithm, into form where each byte is always overwritten, then compilers GCC and clang vectorize it for the AVX2 target:

  for (size_t i=0; i < size; i++)
      bytes[i] = bytes[i] == '/' ? '_' : bytes[i];