[AMD] Use -triton-licm to replace general -licm (#6212)

This picks up #6051
to enable hoisting loop invariant loads in the AMD backend.

Author: antiagainst

third_party/amd/backend/compiler.py

@@ -208,7 +208,7 @@ def make_ttir(mod, metadata, options):
         passes.ttir.add_combine(pm)
         passes.ttir.add_reorder_broadcast(pm)
         passes.common.add_cse(pm)
-        passes.common.add_licm(pm)
+        passes.ttir.add_triton_licm(pm)
         passes.common.add_symbol_dce(pm)
         passes.ttir.add_loop_unroll(pm)
         pm.run(mod)
@@ -234,7 +234,7 @@ def make_ttgir(mod, metadata, options):
 
         passes.ttgpuir.add_fuse_nested_loops(pm)
         passes.common.add_canonicalizer(pm)
-        passes.common.add_licm(pm)
+        passes.ttir.add_triton_licm(pm)
         passes.common.add_canonicalizer(pm)
 
         global_prefetch = int(os.getenv("TRITON_HIP_GLOBAL_PREFETCH", "0"))