[AMD] Enable automatic loop fusion (#5876)

Pick up #5726
for the AMD backend.

Author: antiagainst

python/test/unit/language/test_core.py

@@ -6732,8 +6732,6 @@ def test_tl_range_num_stages(device):
 
 
 def test_tl_range_fuse():
-    if is_hip():
-        pytest.skip("loop fusion is not enabled on AMD")
 
     @triton.jit
     def kernel(ub):

third_party/amd/backend/compiler.py

@@ -232,6 +232,11 @@ def make_ttgir(mod, metadata, options):
         passes.ttgpuir.add_optimize_dot_operands(pm, True)
         amd.passes.ttgpuir.add_hoist_layout_conversions(pm)
 
+        passes.ttgpuir.add_fuse_nested_loops(pm)
+        passes.common.add_canonicalizer(pm)
+        passes.common.add_licm(pm)
+        passes.common.add_canonicalizer(pm)
+
         global_prefetch = int(os.getenv("TRITON_HIP_GLOBAL_PREFETCH", "0"))
         local_prefetch = int(os.getenv("TRITON_HIP_LOCAL_PREFETCH", "0"))