Optimization issue - how do I use normal registers and loop unrolling

AaronNGray · January 6, 2023, 5:23pm

I am trying to generate loop unrolling for a lookahead guard mechanism for a recursive descent parser.

Here is the code :-

gist.github.com

https://gist.github.com/AaronNGray/1bee1d1e494b5c30b2bd3b197710410b

bitset-AND.cpp

#include <cstddef>
#include <initializer_list>

template<size_t BITS>
class bitset
{
public:
    bitset(bool complement = false) : sizeInBits(BITS), sizeOfWord(sizeof(WORD) * 8), sizeInWords((sizeInBits + sizeOfWord - 1) / sizeOfWord), data(NULL) {
        data = new WORD[sizeInWords];
        for (size_t i = 0; i < sizeInWords; ++i)

This file has been truncated. show original

And heres what GodBolt is giving me for Clang, its using xmm, when I really want this to boil down to a 64 bit AND and a JNZ :-

GCC is giving me alright results for this :-

If someone who knows about Clang optimization can have a look at this please ?

Many thanks in advance,

Aaron

AaronNGray · January 6, 2023, 9:57pm

Okay I think I have rationalized things a touch :-

gist.github.com

https://gist.github.com/AaronNGray/498b276ddfb05369e404f937aa091fc0

bitset lookahead-ANDNZ.cpp

#include <cstddef>
#include <initializer_list>

template<size_t BITS>
class bitset
{
public:
    bitset() : sizeInBits(BITS), sizeOfWord(sizeof(WORD) * 8), sizeInWords((BITS + (sizeof(WORD) * 8) - 1) / (sizeof(WORD) * 8)) {
        for (size_t i = 0; i < sizeInWords; ++i)
            data[i] = 0;

This file has been truncated. show original

Both come down to a TEST instruction near enough basically :-

MS Visual Studio C++ is a nightmare still though :-

AaronNGray · January 7, 2023, 12:47am

Solved MS VS C++ on command line, getting simular results, although not as good as Clang or GCC.

I am now looking at proper implementation in C++, and also C.