[llvm] [AMDGPU][MC] Support UC_VERSION_* constants. (PR #95618)

[Fangrui Song via llvm-commits]

MaskRay wrote:

After this patch,
```
llvm-mc -triple=amdgcn -mcpu=gfx1200 -disassemble -show-encoding < gfx12_dasm_sopk.txt
```
will fail when I make the following change (part of efforts to clean up some tech debt in MC)

```
--- i/llvm/lib/MC/MCExpr.cpp
+++ w/llvm/lib/MC/MCExpr.cpp
@@ -534,6 +534,5 @@ bool MCExpr::evaluateAsRelocatableImpl(MCValue &Res, const MCAssembler *Asm,

     // Evaluate recursively if this is a variable.
-    if (Sym.isVariable() && (Kind == MCSymbolRefExpr::VK_None || Layout) &&
-        canExpand(Sym, InSet)) {
+    if (Sym.isVariable() && Layout && canExpand(Sym, InSet)) {
       bool IsMachO = SRE->hasSubsectionsViaSymbols();
       if (Sym.getVariableValue()->evaluateAsRelocatableImpl(Res, Asm, Addrs,
```

In the new code, `MO.getExpr()->evaluateAsAbsolute(Val)` passes nullptr to `MCAssembler *Asm` (we are not generating a relocatable file) and expects MC to expand the variable `UC_VERSION_GFX10`.
This is a behavior that MC wants to migrate away from (`evaluateAsRelocatableImpl` is called with both null Asm and non-null Asm due to tech debt. We want to do less work at parse time and postpone `evaluateAsRelocatableImpl` to object file generation time.

It's better to add extra code the expand the variable manually.


https://github.com/llvm/llvm-project/pull/95618