asm | : running : An x86-64 assembler written in Go | Video Game library

In my opinion, both

Add mavenCentral() in Build Script

Old x86-64 CPUs don't support lzcnt, so rustc/llvm won't emit it by default. (They would execute it as bsr but the behavior is not identical.)

Use -C target-feature=+lzcnt to enable it. Try.

More generally, you may wish to use -C target-cpu=XXX to enable all the features of a specific CPU model. Use rustc --print target-cpus for a list.

In particular, -C target-cpu=native will generate code for the CPU that rustc itself is running on, e.g. if you will run the code on the same machine where you are compiling it.

OK, thank you all, for making me understand the process of DLL importing.

As IInspectable and Remy Lebeau said - the import of DLL requires linking with the LIB. Here is more explanations. Also google - "linking a shared library to executable". It is not important whether it is .so or .dll, the principals are the same.

One other important point before I give a solution.

As Remy Lebeau said: several functions

didn't exist yet (or were introduced shortly before) when BCB5 was released

Fix for makefile

The ARM64 orr immediate instruction takes a bitmask immediate, see Range of immediate values in ARMv8 A64 assembly for an explanation. And GCC has a constraint for an operand of this type: L.

So I would write:

A solution would be if in Termux app you will do next things: (more details here)

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:

Zen2 has 3 cycle latency for vaddpd, 5 cycle latency for vfma...pd. (https://uops.info/).

Your code with 8 accumulators has enough ILP that you'd expect close to two FMA per clock, about 8 per 5 clocks (if there aren't other bottlenecks) which is a bit less than the 10/5 theoretical max.

vaddpd and vmulpd actually run on different ports on Zen2 (unlike Intel), port FP2/3 and FP0/1 respectively, so it can in theory sustain 2/clock vaddpd and vmulpd. Since the latency of the loop-carried dependency is shorter, 8 accumulators are enough to hide the vaddpd latency if scheduling doesn't let one dep chain get behind. (But at least multiplies aren't stealing cycles from it.)

Zen2's front-end is 5 instructions wide (or 6 uops if there are any multi-uop instructions), and it can decode memory-source instructions as a single uop. So it might well be doing 2/clock each multiply and add with the non-FMA version.

If you can unroll by 10 or 12, that might hide enough FMA latency and make it equal to the non-FMA version, but with less power consumption and more SMT-friendly to code running on the other logical core. (10 = 5 x 2 would be just barely enough, which means any scheduling imperfections lose progress on a dep chain which is on the critical path. See Why does mulss take only 3 cycles on Haswell, different from Agner's instruction tables? (Unrolling FP loops with multiple accumulators) for some testing on Intel.)

(By comparison, Intel Skylake runs vaddpd/vmulpd on the same ports with the same latency as vfma...pd, all with 4c latency, 0.5c throughput.)

I didn't look at your code super carefully, but 10 YMM vectors might be a tradeoff between touching two pairs of cache lines vs. touching 5 total lines, which might be worse if a spatial prefetcher tries to complete an aligned pair. Or might be fine. 12 YMM vectors would be three pairs, which should be fine.

Depending on matrix size, out-of-order exec may be able to overlap inner loop dep chains between separate iterations of the outer loop, especially if the loop exit condition can execute sooner and resolve the mispredict (if there is one) while FP work is still in flight. That's an advantage to having fewer total uops for the same work, favouring FMA.

The correct header to use is asm/current.h, do not use asm-generic. This applies to anything under asm really. Headers in the asm-generic folder are provided (as the name suggests) as a "generic" default implementation of macros/functions, then each architecture /arch/xxx has its own asm include folder, where if needed it can define the same macros/functions in an architecture-specific way.

This is done both because it could be actually needed (some archs might have an implementation that is not compatible with the generic one) and for performance since there might be a better and more optimized way of achieving the same result under a specific arch.

Indeed, if we look at how each arch defines get_current() or get_current_thread_info() we can see that some of them (e.g. alpha, spark) keep a reference to the current task in the thread_info struct and keep a pointer to the current thread_info in a register for performance. Others directly keep a pointer to current in a register (e.g. powerpc 32bit), and others define a global per-cpu variable (e.g. x86). On x86 in particular, the thread_info struct doesn't even have a pointer to the current task, it's a very simple 16-byte structure made to fit in a cache line for performance.

Only a few specific GPR instructions have VEX encodings, primarily the BMI1/BMI2 instructions that were added after AVX already existed. See the list in Table 2-28, which has ANDN, BEXTR, BLSI, BLSMSK, BLSR, BZHI, MULX, PDEP, PEXT, RORX, SARX, SHLX, SHRX, as well as the same list in 5.1.16.1. For example, andn's manual entry lists only a VEX encoding, and's manual entry doesn't list any.

So Intel (unfortunately) didn't introduce a brand new three-operand alternate encoding for the entire instruction set. They just introduced a few specific instructions that take three operands and use VEX for it. In some cases these have similar or equivalent functionality to an existing instruction, e.g. SHLX for SHL with a variable count, and so effectively provide a three-operand version of the previous two-operand instruction, but only in those special cases. There are not equivalent instructions across the board.

The "old style" two-operand form remains the only version of the add instruction. However, as fuz points out in comments, lea can be a good way to add two registers and write the result to a third, subject to some restrictions on operand size.

See Using LEA on values that aren't addresses / pointers? for more general things LEA can do, like copy-and-add a constant to a register, or shift-and-add. Compilers already know this and will use lea where appropriate, any time it saves instructions. (Or with some tune options like -mtune=atom for old in-order Atom, will use lea even when they could have used add.)

If more flexible encodings of common integer instructions other than add existed, like and/xor/sub, gcc -O3 -march=skylake would already be using them in its own asm output, without needing inline asm. Or if alternative instructions could get the job done, like lea for add, would be doing that, so it makes sense to look at compiler output to see what tricks it knows. Trying it yourself would make more sense as something to play around with in a stand-alone .s file that just makes an exit system call, or just to single-step, removing the complexity of using inline asm. (GAS by default doesn't restrict instruction-sets. gcc -march=skylake doesn't pass that on to the assembler, as.)

In your inline asm, your c operand should be to output-only: =r instead of +r. The old value is overwritten, so there's no need to tell the compiler to produce it as an input. (Like you said, you want c = a+b not c += a+b.)

Using a single lea as the asm template means you don't need a =&r early-clobber output, because your asm will read all its inputs before writing that output. In your case, having it as an input/output was probably stopping the compiler from choosing the same register as one of the inputs, which could have broken with mov; add.