Arm64Emitter: Improve MOVI2R

[JosJuice]

More or less a complete rewrite of the function which aims to be equally good or better for each given input, without relying on special cases like the old implementation did.

In particular, we now have more extensive support for MOVN, as mentioned in a TODO comment.

[JosJuice]

Now that PR #9423 exists, ideally this PR should be merged after that one.

[JosJuice]

Added ORR support, made us prefer ADRP+ADD over ADR+MOVK and ADRP+MOVK, and fixed a bug in my ADRP code.

[degasus]

If I calculated correctly, this code will call malloc 8 times per MOVI2R. I fear this might have a significat time hit in code compilation.

[JosJuice]

What do you think about replacing std::vector<Part> with std::array<Part, 4> plus a size_t to keep track of the size?

[degasus]

I assume the std::array will make this a lot faster. I'm also fine with the vector if you tell me that it has no large impact in your profile ;)

[JosJuice]

Done.

[leoetlino]

Changes LGTM. The unit tests still pass, I hope?

[JosJuice]

Yes, the PowerPC tests pass. (The VertexLoader tests still aren't passing, like in master, but that has no relation to this PR.)

[dovart]

@JosJuice not building on Arm64

../Source/Core/Common/Arm64Emitter.cpp: In member function ‘void Arm64Gen::ARM64XEmitter::MOVI2RImpl(Arm64Gen::ARM64Reg, T)’:
../Source/Core/Common/Arm64Emitter.cpp:2039:5: internal compiler error: Segmentation fault
     MOVZ,
     ^~~~
0x7f99269d23 __libc_start_main
	../csu/libc-start.c:308

[degasus]

@dovart This is an internal compiler error. So which compiler are you using? Which version of the compiler is it? Is this for android, windows or linux?

[dovart]

@degasus gcc version 8.3.0 (Debian 8.3.0-6)
platform: pi

[dovart]

@degasus was building fine before the commits from 2 days ago.

[JosJuice]

I wonder if it works fine on newer versions of GCC... If so, maybe it's time to simply drop support for GCC 8. (See PR #9294)

[merryhime]

I can reproduce the above ICE on gcc (Debian 8.3.0-6) 8.3.0 on a desktop AArch64 Debian buster system.
Installed gcc-10 (most recent package available via testing).
Unfortunately this ICE remains on the latest version of GCC available via debian's package repository (gcc (Debian 10.2.1-6) 10.2.1 20210110).
Will submit a GCC bugreport once I get an account on their bugzilla.

---

Reduced to minimal testcase:

Required command line:
g++ -Wshadow a.cpp

a.cpp:

class ARM64XEmitter
{
  void MOVZ();

  template <typename T>
  void MOVI2RImpl(T imm);
};

void ARM64XEmitter::MOVZ()
{
}

template <typename T>
void ARM64XEmitter::MOVI2RImpl(T imm)
{
  enum class Approach
  {
    MOVZ,
  };
}

Bug in shadow detection. Having the member function be templated is required for ICE to occur.

---

Submitted GCC bug. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99120