diff --git a/loopy/codegen/loop.py b/loopy/codegen/loop.py
index 0110a06095fa0bd690045f050136027d7bed3a28..afb23fe4e3af4b9ac8191ed5d7cbda14bc287dd9 100644
--- a/loopy/codegen/loop.py
+++ b/loopy/codegen/loop.py
@@ -410,17 +410,21 @@ def generate_sequential_loop_dim_code(codegen_state, sched_index):
 
         # {{{ find implemented loop, build inner code
 
-        from loopy.symbolic import pw_aff_to_pw_aff_implemented_by_expr
-        impl_lbound = pw_aff_to_pw_aff_implemented_by_expr(lbound)
-        impl_ubound = pw_aff_to_pw_aff_implemented_by_expr(ubound)
+        loop_nonempty_params = lbound.params() & ubound.params()
+        has_nonempty_check = False
+
+        if not loop_nonempty_params.plain_is_universe():
+            has_nonempty_check = True
+            from loopy.symbolic import set_to_cond_expr
+            loop_nonempty_cond = set_to_cond_expr(loop_nonempty_params)
 
         # impl_loop may be overapproximated
         from loopy.isl_helpers import make_loop_bounds_from_pwaffs
         impl_loop = make_loop_bounds_from_pwaffs(
                 dom_and_slab.space,
                 loop_iname,
-                impl_lbound,
-                impl_ubound)
+                lbound,
+                ubound)
 
         for moved_iname in moved_inames:
             # move moved_iname to 'set' dim_type in impl_loop
@@ -446,6 +450,16 @@ def generate_sequential_loop_dim_code(codegen_state, sched_index):
 
         from loopy.symbolic import pw_aff_to_expr
 
+        def add_nonempty_check_maybe(result):
+            if has_nonempty_check:
+                from loopy.codegen.result import wrap_in_if
+                return wrap_in_if(
+                        codegen_state,
+                        [loop_nonempty_cond],
+                        result)
+            else:
+                return result
+
         if ubound.is_equal(lbound):
             # single-trip, generate just a variable assignment, not a loop
             inner = merge_codegen_results(codegen_state, [
@@ -458,10 +472,11 @@ def generate_sequential_loop_dim_code(codegen_state, sched_index):
                 inner,
                 ])
             result.append(
+                add_nonempty_check_maybe(
                     inner.with_new_ast(
                         codegen_state,
                         astb.ast_block_scope_class(
-                            inner.current_ast(codegen_state))))
+                            inner.current_ast(codegen_state)))))
 
         else:
             inner_ast = inner.current_ast(codegen_state)
@@ -469,13 +484,14 @@ def generate_sequential_loop_dim_code(codegen_state, sched_index):
             from loopy.isl_helpers import simplify_pw_aff
 
             result.append(
-                inner.with_new_ast(
-                    codegen_state,
-                    astb.emit_sequential_loop(
-                        codegen_state, loop_iname, kernel.index_dtype,
-                        pw_aff_to_expr(simplify_pw_aff(lbound, kernel.assumptions)),
-                        pw_aff_to_expr(simplify_pw_aff(ubound, kernel.assumptions)),
-                        inner_ast)))
+                add_nonempty_check_maybe(
+                    inner.with_new_ast(
+                        codegen_state,
+                        astb.emit_sequential_loop(
+                            codegen_state, loop_iname, kernel.index_dtype,
+                            pw_aff_to_expr(simplify_pw_aff(lbound, kernel.assumptions)),
+                            pw_aff_to_expr(simplify_pw_aff(ubound, kernel.assumptions)),
+                            inner_ast))))
 
     return merge_codegen_results(codegen_state, result)
 
diff --git a/loopy/symbolic.py b/loopy/symbolic.py
index f1a494f30d469511817d204c0476ff79abe00e3b..3779e37d83e196260ad60bb3f8baa09a285bcecc 100644
--- a/loopy/symbolic.py
+++ b/loopy/symbolic.py
@@ -1190,23 +1190,6 @@ def pw_aff_to_expr(pw_aff, int_ok=False):
 
     return expr
 
-
-def pw_aff_to_pw_aff_implemented_by_expr(pw_aff):
-    pieces = pw_aff.get_pieces()
-
-    rest = isl.Set.universe(pw_aff.space.params())
-    aff_set, aff = pieces[0]
-    impl_pw_aff = isl.PwAff.alloc(aff_set, aff)
-    rest = rest.intersect_params(aff_set.complement())
-
-    for aff_set, aff in pieces[1:-1]:
-        impl_pw_aff = impl_pw_aff.union_max(
-            isl.PwAff.alloc(aff_set, aff))
-        rest = rest.intersect_params(aff_set.complement())
-
-    _, aff = pieces[-1]
-    return impl_pw_aff.union_max(isl.PwAff.alloc(rest, aff)).coalesce()
-
 # }}}
 
 
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 21db62610f3a3160bcc3069c3e480e85cc4712f8..c1739e151e69ad7b63fd1c0a37bc392f077cb12e 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -2298,6 +2298,45 @@ def test_struct_assignment(ctx_factory):
     knl(queue, N=200)
 
 
+def test_hoisting_of_conditionals_for_nonempty_loop():
+    # This checks that conditionals that are redundant when the loop is
+    # non-empty aren't generated by the bounds check code inside the loop.
+
+    dtype = np.float32
+
+    # group size
+    by = 16
+    bx = 24
+
+    # define kernel
+    knl = lp.make_kernel(
+            "{[i,k,j]: 0<=i<n and 0<=k<m and 0<=j<l}",
+            [
+                "c[i, j] = sum(k, a[i, k]*b[k, j])"
+            ], name="nonsq_mm", assumptions="n,m,l >= 1")
+
+    knl = lp.add_dtypes(knl, dict(a=dtype, b=dtype, c=dtype))
+
+    # transform
+    knl = lp.split_iname(knl, "i", by, outer_tag="g.0", inner_tag="l.1")
+    knl = lp.split_iname(knl, "j", bx, outer_tag="g.1", inner_tag="l.0")
+    knl = lp.split_iname(knl, "k", by)
+    knl = lp.extract_subst(knl, "a_acc", "a[i1,i2]", parameters="i1, i2")
+    knl = lp.extract_subst(knl, "b_acc", "b[i1,i2]", parameters="i1, i2")
+    knl = lp.precompute(knl, "a_acc", "k_inner,i_inner")
+    knl = lp.precompute(knl, "b_acc", "j_inner,k_inner")
+
+    cgr = lp.generate_code_v2(knl)
+
+    expected_substr = (
+            "for (int k_inner = 0; k_inner <= (15 + -1 * lid(0) >= 0 && -17 + m"
+            " + -16 * k_outer >= 0 ? 15 : -1 + m + -16 * k_outer); ++k_inner)"
+            "\n"
+            "          acc_k_outer_k_inner = acc_k_outer_k_inner + ")
+
+    assert expected_substr in cgr.device_code()
+
+
 def test_inames_conditional_generation(ctx_factory):
     ctx = ctx_factory()
     knl = lp.make_kernel(