[llvm] [RegAllocFast] fold foldable inline asm (PR #74344)

[Matthias Braun via llvm-commits]

MatzeB wrote:

> Is that something you'd imagine occurring within fastregalloc?

> This PR essentially is a 2 pass algorithm; it does one pass for foldable operands first, then defers to existing machinery for the non-foldable operands. It's not clear from your suggestion of doing that in the opposite order what the improvement would be (if it's even feasible); can you provide more thoughts on what you have in mind?

Yes I was referring to the loop prefixed with `// Allocate virtreg uses and insert reloads as necessary.` and was wondering if we could replace this with something like this pseudo code:

```
// Allocate virtreg uses and insert reloads as necessary.
if (isInlineAsm(MI)) {
  assignInlineAsmUses(MI...);
} else {
  ... original loop to assign uses ...
}

...

void assignInlineAsmUses(MI) {
   // We do two passes to assign the non-foldable operands first, as they require an assigned physical register while
   // foldable operands can always be spilled if there are no registers left.
   // first pass for non-foldable operands
   ... original loop to assign uses, but with added `if (isFoldableMemOp(MO)) continue;` ...
   // 2nd pass for foldable operands
   ... original loop to assign uses, but with added `if (!isFoldableMemOp(MO)) continue;` ...
}
```

https://github.com/llvm/llvm-project/pull/74344