Sy BrandNov 28, 2022

I co-wrote a tour of some recent MSVC backend improvements!

Have a look for some before-and-after assembly samples and more details.

https://devblogs.microsoft.com/cppblog/a-tour-of-4-msvc-backend-improvements/

#cpp

A Tour of 4 MSVC Backend Improvements

This blog post presents some of the optimizations the backend team has implemented for Visual Studio 2022.

C++ Team Blog
Show threadSy BrandNov 28, 2022

One is a bit tracker which can recognise custom byteswap functions and optimise down to bswap

Before > After

#cpp

Show threadSy BrandNov 28, 2022

Another is loop unswitching, which tries to avoid unnecessary branching based on invariant conditions in loops.

It basically does the equivalent of translating the 1st C++ sample into the 2nd one

#cpp

Show threadSy Brand

Next is better codegen for min/max chains.

Given the 1st C++ sample, the compiler used to generate what's in the 2nd pic. Now it generates the 3rd.

#cpp

Nov 28, 2022 at 4:17pmWeb
Show threadFCLCNov 28, 2022
@TartanLlama Oh, damn, that's WAYYY better. is this SSE only mode? and does this scale to the various AVX variants?
Show threadSy BrandNov 28, 2022
@fclc this is with /O2 and /fp:fast, which is required
Show threadFélixNov 28, 2022
@TartanLlama 🥵