GCC

[ghost]

Now compiles on GCC and Clang.

[ArtemisX64]

Can confirm using gcc also fixes the final linking error

[ArtemisX64]

Weird but it no longer compiles on clang(clang 12)
Oops as expected. Won't compile on Windows too. (Checked in actions on my repo)

[ArtemisX64]

It will work fine in clang14(libstdc++12)

[Tachi107]

Thanks for working on this! I've left a few comments below.

[Tachi107]

What kind of UB did you encounter?

…

-- OpenPGP key: 66DE F152 8299 0C21 99EF A801 A8A1 28A8 AB1C EE49

[ArtemisX64]

What kind of UB did you encounter?
…
-- OpenPGP key: 66DE F152 8299 0C21 99EF A801 A8A1 28A8 AB1C EE49

Can't recollect. But I think it will be fine. (As long as it receives uint64_t as a param)

[Tachi107]

What kind of UB did you encounter?

Can't recollect. But I think it will be fine. (As long as it receives uint64_t as a param)

Maybe the issue was that the parameter type was uint64 instead of uint64_t, and numeric_limits<uint64>​::​digits doesn't work on some platforms?

[Crementif]

LGTM

[bitscher]

The addition of the -mavx2 flag is causing issues on pre-Haswell CPUs #144

[ghost]

-mavx2 is required for all the _mm256_... functions.
-mssse3 is for _mm_shuffle_epi8

Other instructions used in Cemu:

• sse _mm_prefetch, _mm_pause, _mm_setcsr
• sse2 _mm_mfence, _mm_set_epi8, _mm_set_epi16, _mm_loadu_si128, _mm_store_si128
• sse4.1 _mm_min_epu16, _mm_max_epu16

So it may actually be a good idea to add additional flags.

[Tachi107]

What about adding -march=x86-64-v2 and trying to stick to the x86_64-v2 psABI?

Edit: nevermind, AVX2 is part of v3

[Exzap]

Both clang and gcc have a way to selectively enable CPU extensions for individual functions. Here is an example. It's better to do it this way than to set -mavx2 or similar globally as Cemu doesn't have a hard requirement on any of the optional CPU extensions.

[ghost]

Both clang and gcc have a way to selectively enable CPU extensions for individual functions. Here is an example. It's better to do it this way than to set -mavx2 or similar globally

I agree, I can make a pr changing this if you wish but I don't think it would help insofar as compiling Cemu for pre-haswell CPUs on GCC.
From https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#Common-Function-Attributes

For instance, on an x86, you could declare one function with the target("sse4.1,arch=core2") attribute and another with target("sse4a,arch=amdfam10"). This is equivalent to compiling the first function with -msse4.1 and -march=core2 options, and the second function with -msse4a and -march=amdfam10 options. It is up to you to make sure that a function is only invoked on a machine that supports the particular ISA it is compiled for

[Exzap]

I agree, I can make a pr changing this if you wish

It would be appreciated.

I don't think it would help insofar as compiling Cemu for pre-haswell CPUs on GCC.

Why not? As it says in your quote from the spec, you have to manually make sure the CPU extensions are supported at runtime before running the function with the intrinsics. This is something we already do. Optionally running the hand-optimized functions if the CPU extensions are supported and otherwise falling back to a generic implementation that can run on any CPU. The code for handling GPU drawcall indices is a good example. On Windows this is all pretty well tested and people run Cemu with the most ancient of CPUs.

[ghost]

Alright, I didn't realise you already checked this, thanks.

[ghost]

Here you go #152

[ghost]

I must have missed these clang specific pragmas when I made #80. Sorry about that.

Cemu/src/Cafe/HW/Latte/Core/LatteIndices.cpp

Line 288 in 68fa5b3

#pragma clang attribute push (__attribute__((target("avx2"))), apply_to=function)