[llvm] [AMDGPU] Work around s_getpc_b64 zero extending on GFX12 (PR #78186)

[Matt Arsenault via llvm-commits]

================
@@ -1810,6 +1810,8 @@ def int_amdgcn_s_setreg :
 // not cross a 4Gb address boundary. Use for any other purpose may not
 // produce the desired results as optimizations may cause code movement,
 // especially as we explicitly use IntrNoMem to allow optimizations.
+// This intrinsic always returns PC sign-extended from 48 bits even if the
+// s_getpc_b64 instruction returns a zero-extended value.
----------------
arsenm wrote:

Maybe add this to AMDGPUUsage's intrinsic documentation?

https://github.com/llvm/llvm-project/pull/78186