https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88152
Bug ID: 88152
Summary: optimize SSE & AVX char compares with subsequent
movmskb
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: kretz at kde dot org
Target Milestone: ---
Target: x86_64-*-*, i?86-*-*
Testcase (https://godbolt.org/z/YNPZyf):
#include <x86intrin.h>
template <typename T, size_t N>
using V [[gnu::vector_size(N)]] = T;
// the following should be optimized to:
// vpxor %xmm1, %xmm1, %xmm1
// vpcmpgtb %[xy]mm0, %[xy]mm1, %[xy]mm0
// ret
auto cmp0(V<unsigned char, 16> a) { return a > 0x7f; }
auto cmp0(V<unsigned char, 32> a) { return a > 0x7f; }
auto cmp1(V<unsigned char, 16> a) { return a >= 0x80; }
auto cmp1(V<unsigned char, 32> a) { return a >= 0x80; }
auto cmp0(V< signed char, 16> a) { return a < 0; }
auto cmp0(V< signed char, 32> a) { return a < 0; }
auto cmp1(V< signed char, 16> a) { return a <= -1; }
auto cmp1(V< signed char, 32> a) { return a <= -1; }
auto cmp0(V< char, 16> a) { return a < 0; }
auto cmp0(V< char, 32> a) { return a < 0; }
auto cmp1(V< char, 16> a) { return a <= -1; }
auto cmp1(V< char, 32> a) { return a <= -1; }
// the following should be optimized to:
// vpmovmskb %[xy]mm0, %eax
// ret
int f0(V<unsigned char, 16> a) {
return _mm_movemask_epi8 (reinterpret_cast<__m128i>(a > 0x7f));
}
int f0(V<unsigned char, 32> a) {
return _mm256_movemask_epi8(reinterpret_cast<__m256i>(a > 0x7f));
}
int f1(V<unsigned char, 16> a) {
return _mm_movemask_epi8 (reinterpret_cast<__m128i>(a >= 0x80));
}
int f1(V<unsigned char, 32> a) {
return _mm256_movemask_epi8(reinterpret_cast<__m256i>(a >= 0x80));
}
int f0(V< signed char, 16> a) {
return _mm_movemask_epi8 (reinterpret_cast<__m128i>(a < 0));
}
int f0(V< signed char, 32> a) {
return _mm256_movemask_epi8(reinterpret_cast<__m256i>(a < 0));
}
int f1(V< signed char, 16> a) {
return _mm_movemask_epi8 (reinterpret_cast<__m128i>(a <= -1));
}
int f1(V< signed char, 32> a) {
return _mm256_movemask_epi8(reinterpret_cast<__m256i>(a <= -1));
}
int f0(V< char, 16> a) {
return _mm_movemask_epi8 (reinterpret_cast<__m128i>(a < 0));
}
int f0(V< char, 32> a) {
return _mm256_movemask_epi8(reinterpret_cast<__m256i>(a < 0));
}
int f1(V< char, 16> a) {
return _mm_movemask_epi8 (reinterpret_cast<__m128i>(a <= -1));
}
int f1(V< char, 32> a) {
return _mm256_movemask_epi8(reinterpret_cast<__m256i>(a <= -1));
}
Compile with `-O2 -mavx2` (the same is relevant with -msse2 if you remove the
AVX overloads).
Motivation:
This pattern is relevant for vectorized UTF-8 decoding, where all bytes with
MSB == 0 can simply be zero extended to UTF-16/32. Such code could just skip
the compare and call movemask directly on `a`. However:
std::experimental::simd doesn't (and any other general purpose SIMD abstraction
shouldn't) expose a "munch sign bits into bitmask integer" function. Such a
function is too ISA specific.
In the interest of making code readable (and thus maintainable) I strongly
believe it should read: `n_ascii_chars = find_first_set(a > 0x7f)` while still
getting the optimization.
Similar test cases can be constructed for movmskp[sd] after 32/64 bit integer
compares.
