o2 | 2D Game Engine with visual WYSIWYG editor | Game Engine library

update: GCC will do this for you with -ftracer (https://godbolt.org/z/es7a3bEPv), optimizing away the swap. (That's only enabled manually or as part of -fprofile-use, not at -O3, so it's probably not a good idea to use all the time without PGO, potentially bloating machine code in cold functions / code-paths.)

Doing it manually in the source (Godbolt):

• Nobody found any undefined behavior in the bug repro code.
• At least some of you were able to reproduce the undesired behavior.

So I filed a bug report against Visual Studio 2019.

The Microsoft team confirmed the problem.

However, unfortunately it seems like Visual Studio 2019 will not receive a bug fix because Visual Studio 2022 seemingly does not have the bug. Apparently, the most recent version not having that particular bug is good enough for Microsoft's quality standards.

I find this disappointing because I think that the correctness of a compiler is essential and Visual Studio 2022 has just been released with new features and therefore probably contains new bugs. So there is no real "stable version" (one is cutting edge, the other one doesn't get bug fixes). But I guess we have to live with that or choose a different, more stable compiler.

The option is however considered, by some, a bit outdated by now, but in this case, it's finding a real problem.

Currently (2022-02-05), CRAN builds R binaries for Apple silicon using Apple clang (from Command Line Tools for Xcode 12.4) and an experimental build of gfortran.

If you obtain R from CRAN (i.e., here), then you need to replicate CRAN's compiler setup on your system before building R packages that contain C/C++/Fortran code from their sources (and before using Rcpp, etc.). This requirement ensures that your package builds are compatible with R itself.

A further complication is the fact that Apple clang doesn't support OpenMP, so you need to do even more work to compile programs that make use of multithreading. You could circumvent the issue by building R itself and all R packages from sources with LLVM clang, which does support OpenMP, but this approach is onerous and "for experts only". There is another approach that has been tested by a few people, including Simon Urbanek, the maintainer of R for macOS. It is experimental and also "for experts only", but seems to work on my machine and is simpler than trying to build R yourself.

Warning: These instructions come with no warranty and could break at any time. They assume some level of familiarity with C/C++/Fortran program compilation, Makefile syntax, and Unix shells. As usual, sudo at your own risk.

I will try to address compilers and OpenMP support at the same time. I am going to assume that you are starting from nothing. Feel free to skip steps you've already taken, though you might find a fresh start helpful.

I've tested these instructions on a machine running Big Sur, and at least one person has tested them on a machine running Monterey. I would be glad to hear from others.

1. Download an R binary from CRAN here and install. Be sure to select the binary built for Apple silicon.

2. Run

(gcc -O3 enables -ftree-vectorize and a few other options not included by -O2, e.g. if-conversion to branchless cmov, which is another way -O3 can hurt with data patterns GCC didn't expect. By comparison, Clang enables auto-vectorization even at -O2, although some of its optimizations are still only on at -O3.)

It's doing 64-bit loads (and branching to store or not) on pairs of ints. This means, if we swapped the last iteration, this load comes half from that store, half from fresh memory, so we get a store-forwarding stall after every swap. But bubble sort often has long chains of swapping every iteration as an element bubbles far, so this is really bad.

(Bubble sort is bad in general, especially if implemented naively without keeping the previous iteration's second element around in a register. It can be interesting to analyze the asm details of exactly why it sucks, so it is fair enough for wanting to try.)

Anyway, this is pretty clearly an anti-optimization you should report on GCC Bugzilla with the "missed-optimization" keyword. Scalar loads are cheap, and store-forwarding stalls are costly. (Can modern x86 implementations store-forward from more than one prior store? no, nor can microarchitectures other than in-order Atom efficiently load when it partially overlaps with one previous store, and partially from data that has to come from the L1d cache.)

Even better would be to keep buf[x+1] in a register and use it as buf[x] in the next iteration, avoiding a store and load. (Like good hand-written asm bubble sort examples, a few of which exist on Stack Overflow.)

If it wasn't for the store-forwarding stalls (which AFAIK GCC doesn't know about in its cost model), this strategy might be about break-even. SSE 4.1 for a branchless pmind / pmaxd comparator might be interesting, but that would mean always storing and the C source doesn't do that.

If this strategy of double-width load had any merit, it would be better implemented with pure integer on a 64-bit machine like x86-64, where you can operate on just the low 32 bits with garbage (or valuable data) in the upper half. E.g.,

1. pkg install proot
2. pkg install proot-distro
3. proot-distro install debian
4. proot-distro login debian

After that you should be logged in a Debian environment, and you can install almost any Arm packages available on debian repositories.

For example you should be able to install this Cortex compiler:

This means that the optimized version can be executed with only 2 cycle of latency while the first one takes 3 cycle of latency (not taking into account load/store instructions that are the same). Moreover, the second version can be better pipelined since the two instructions can be executed for two different input data in parallel thanks to a superscalar out-of-order execution. Note that two loads can be executed in parallel too although only one store can be executed in parallel per cycle. This means that the execution is bounded by the throughput of store instructions. Overall, only 1 value can only computed per cycle. AFAIK, recent Intel Icelake processors can do two stores in parallel like new AMD Ryzen processors. The second one is expected to be as fast or possibly faster on the chosen use-case (Intel Skylake processors). It should be significantly faster on very recent x86-64 processors.

Note that the lea instruction is very fast because the multiply-add is done on a dedicated CPU unit (hard-wired shifters) and it only supports some specific constant for the multiplication (supported factors are 1, 2, 4 and 8, which mean that lea can be used to multiply an integer by the constants 2, 3, 4, 5, 8 and 9). This is why lea is faster than imul/mul.

I can reproduce the slower execution with -O2 using GCC 11.2 (on Linux with a i5-9600KF processor).

The main source of source of slowdown comes from the higher number of micro-operations (uops) to be executed in the -O2 version certainly combined with the saturation of some execution ports certainly due to a bad micro-operation scheduling.

Here is the assembly of the loop with -Os: