[llvm] [AArch64] Disable coalescing of SUBREG_TO_REG with subreg liveness. (PR #174185)

[Sander de Smalen via llvm-commits]

================
@@ -214,8 +218,10 @@ define void @t17(i64 %a) {
 define i8 @LdOffset_i8(ptr %a)  {
 ; CHECK-LABEL: LdOffset_i8:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    add x8, x0, #253, lsl #12 // =1036288
-; CHECK-NEXT:    ldrb w0, [x8, #3704]
----------------
sdesmalen-arm wrote:

The SUBREG_TO_REG gets replaced by a KILL instruction, which is not handled by the AArch64 load-store optimization pass that tries to fold the immediate into the operation.
```
renamable $w8 = MOVi32imm 1039992
renamable $x8 = KILL killed renamable $w8
renamable $w0 = LDRBBroX killed renamable $x0, killed renamable $x8, 0, 0 :: (load (s8) from %ir.arrayidx)
RET_ReallyLR implicit $w0
```

https://github.com/llvm/llvm-project/pull/174185