Implement black_box using intrinsic

[nbdd0121]

Introduce black_box intrinsic, as suggested in #87590 (comment).

This is still codegenned as empty inline assembly for LLVM. For MIR interpretation and cranelift it's treated as identity.

cc @Amanieu as this is related to inline assembly
cc @bjorn3 for rustc_codegen_cranelift changes
cc @RalfJung as this affects MIRI

r? @nagisa I suppose

[rust-highfive]

Some changes occured to rustc_codegen_cranelift

cc @bjorn3

Some changes occured to the CTFE / Miri engine

cc @rust-lang/miri

[nagisa]

Looks pretty nice. And in the future we have freedom to implement this in a better or otherwise backend specific manner better.

[nagisa]

@bors delegate=nbdd0121

(feel free to r=nagisa once the unwrap is bug!ified)

[bors]

✌️ @nbdd0121 can now approve this pull request

[nbdd0121]

@bors r=nagisa

[bors]

📌 Commit 3cf2a69 has been approved by nagisa

[RalfJung]

Longer-term we might even want to directly reexport the intrinsic instead of using a wrapper function... but that's blocked on the fn ptr reification PR.

[nbdd0121]

I guess that's also blocked by path component stability

[RalfJung]

Well, we already reexport transmute and probably will reexport copy/copy_nonoverlapping soon (we did but it was reverted due to the fn ptr reification problem).

So, while the lack of path component stability is a problem, it's not necessarily a blocker.

Anyway, even the exported function is currently still unstable; this is a discussion to be had at stabilization time.

[nbdd0121]

@bors r=nagisa

[bors]

📌 Commit f98d540 has been approved by nagisa

[JohnTitor]

Possibly failed in rollup: #87939 (comment)
@bors r-

[nbdd0121]

failures:

---- [ui] ui/sanitize/memory.rs stdout ----

error: error pattern ' in the stack frame of function 'random'' not found!
command: "/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/memory/a"
stdout:
------------------------------------------


------------------------------------------
stderr:
------------------------------------------
==10866==WARNING: MemorySanitizer: use-of-uninitialized-value
    #0 0x5565d1c6fb44  (/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/memory/a+0x67b44)
    #1 0x7f93163a10b2  (/lib/x86_64-linux-gnu/libc.so.6+0x270b2)
    #2 0x5565d1c111cd  (/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/memory/a+0x91cd)

  Uninitialized value was created by an allocation of 'r.i' in the stack frame of function 'main'
    #0 0x5565d1c6fa40  (/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/memory/a+0x67a40)

SUMMARY: MemorySanitizer: use-of-uninitialized-value (/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/memory/a+0x67b44) 

------------------------------------------

It seems to me that the codegen is better with the intrinsic approach and a memcpy is eliminated from random, so the place where uninit values end up getting used becomes main.

Previously it's like (expressed in MIR):

fn black_box(_1: T) -> T {
    let mut _0: T; // return place
    let mut _2: &mut T;

	bb0: {
	      _2 = &mut _1;
		  llvm_asm!(""::"r"(_2):"memory":"volatile");
	      _0 = move _1; // LLVM can't optimize out
		  return;
	}
}

The intrinsic approach it's codegenned more like:

fn black_box(_1: T) -> T {
    let mut _0: T; // return place
    let mut _2: &mut T;

	bb0: {
		  _0 = move _1; // LLVM can optimize out
	      _2 = &mut _0;
		  llvm_asm!(""::"r"(_2):"memory":"volatile");
		  return;
	}
}

I think it'd be sufficient just to change the pattern so that "main" is also matched?

[nbdd0121]

@nagisa A sanitize test needs to be fixed (last commit). PTAL

[nagisa]

Please squash the commits. r=me once its done.

[nbdd0121]

@bors r=nagisa

[bors]

📌 Commit 1fb1643 has been approved by nagisa

[bors]

⌛ Testing commit 1fb1643 with merge 0fa3190...

[bors]

☀️ Test successful - checks-actions
Approved by: nagisa
Pushing 0fa3190 to master...