Day 22 of Advent of Compiler Optimisations!
Comparing a string_view against "ABCDEFG" should call memcmp, right? Watch what Clang actually generates — no function call at all, just a handful of inline instructions using some rather cunning tricks. How does it compare 7 bytes so efficiently when they don't fit in a single register?
Read more: https://xania.org/202512/22-memory-cunningness
Watch: https://youtu.be/kXmqwJoaapg
@mattgodbolt WTF. I'm really disappointed that Clang is not smart enough to do
matt godbolt
Riley S. Faelant1:
cmp rdi, 1 ; is length 1?
jne .LBB0_1 ; if not 1, goto "return false"
cmp byte ptr [rsi], 65 ; is the byte 65 ('A')?
.LBB0_1:
sete al ; set result to 0 or 1 accordingly
ret ; return
@mattgodbolt my first thought was to just read 8 bytes into one register and mask off the eighth before or after the compare. then i remembered that the eighth might not even be mapped because the start address might be unaligned.
that got me thinking: will there be something about alignment coming up?
@mattgodbolt i've heard of penalties for unaligned accesses at least for some SIMD stuff on x86 and on older ARMs. there are also architectures (ppc?) where you can only access memory in larger aligned words so the compiler needs to do shifting for individual bytes.
i tried to produce something with alignas but had no luck in that regard: https://aoco.compiler-explorer.com/z/jWWvnbxed
please also see how the hover thing for the string in the .quad is shown in reverse, you seem to account for little-endian architectures, but here (on s390x) it is not needed.
using namespace std::literals; template <std::size_t len, std::size_t align> struct alignas(align) str_arr { char a[len]; }; bool t(str_arr<7,1> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCDEFG"sv;} bool t(str_arr<8,1> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCDEFGH"sv;} bool t(str_arr<7,4> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCDEFG"sv;} bool t(str_arr<8,4> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCDEFGH"sv;} bool t(str_arr<8,8> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCDEFGH"sv;} bool t(str_arr<4,4> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCD"sv;} bool t(str_arr<4,1> const * a) { return std::string_view(a->a, sizeof(a->a)) == "ABCD"sv;}
Xarn
Timo Weingärtnerhtonl/ntohl with a loop in a templated function. it works great with clang, but gcc only sees it in one direction, not the other: https://aoco.compiler-explorer.com/z/Wc1PGPTGEtemplate<typename T> auto host_to_network (T value) noexcept { static_assert(std::is_unsigned_v<T>, "shift is only supported on unsigned"); std::array<char, sizeof(T)> ret; for (auto & byte : ret | std::views::reverse) { byte = static_cast<char>(value & 0xff); if constexpr (sizeof(T) > 1) value >>= 8; } return ret; } template<typename T> auto network_to_host (std::array<char, sizeof(T)> const buf) noexcept { static_assert(std::is_unsigned_v<T>, "shift is only supported on unsigned"); T ret{0}; for (auto const byte : buf) { if constexpr (sizeof(T) > 1) ret <<= 8; ret |= static_cast<uint8_t>(byte); } return ret; } template auto host_to_network<uint8_t> (uint8_t value) noexcept; template auto network_to_host<uint8_t> (std::array<char, sizeof(uint8_t)> const buf) noexcept; template auto host_to_network<uint16_t> (uint16_t value) noexcept; template auto network_to_host<uint16_t> (std::array<char, sizeof(uint16_t)> const buf) noexcept; template auto host_to_network<uint32_t> (uint32_t value) noexcept; template auto network_to_host<uint32_t> (std::array<char, sizeof(uint32_t)> const buf) noexcept; template auto host_to_network<uint64_t> (uint64_t value) noexcept; template auto network_to_host<uint64_t> (std::array<char, sizeof(uint64_t)> const buf) noexcept;