[inductor] share cse cache during vectorized indirect load (#124597)

Fix #123502

`swap_buffer` in not needed in vectorized indirect load, remove it to share cse buffer.
```
auto tmp8 =
[&]
{
    __at_align__ std::array<int64_t, 16> tmpbuf;
    tmp7.store(tmpbuf.data());
    return tmpbuf;
}
()
;
//
// other codes
//
// also store tmp7 here (redundant tmp16)
auto tmp16 =
[&]
{
    __at_align__ std::array<int64_t, 16> tmpbuf;
    tmp7.store(tmpbuf.data());
    return tmpbuf;
}
()
;
```

Pull Request resolved: #124597
Approved by: https://github.com/jgong5, https://github.com/jansel

test/inductor/test_cpu_repro.py

@@ -3683,6 +3683,39 @@ def forward(self, x):
         x = torch.randn(1, 4, 2, 2)
         self.common(fn, (x,))
 
+    @requires_vectorization
+    def test_vec_indirect_load_cse_cache(self):
+        # https://github.com/pytorch/pytorch/issues/123502
+        from math import inf
+
+        def fn(arg0_1):
+            full_default = torch.ops.aten.full.default([209985], 1)
+            select = torch.ops.aten.select.int(arg0_1, 0, 0)
+            select_1 = torch.ops.aten.select.int(arg0_1, 0, 1)
+            view = torch.ops.aten.reshape.default(select_1, [-1])
+            expand = torch.ops.aten.expand.default(view, [209985])
+            full_default_1 = torch.ops.aten.full.default([10000], 0)
+            scatter_add = torch.ops.aten.scatter_add.default(
+                full_default_1, 0, expand, full_default
+            )
+            pow_1 = torch.ops.aten.pow.Tensor_Scalar(scatter_add, -0.5)
+            eq = torch.ops.aten.eq.Scalar(pow_1, inf)
+            full_default_2 = torch.ops.aten.full.default([], 0.0)
+            where = torch.ops.aten.where.self(eq, full_default_2, pow_1)
+            index = torch.ops.aten.index.Tensor(where, [select])
+            index_1 = torch.ops.aten.index.Tensor(where, [select_1])
+            mul_1 = torch.ops.aten.mul.Tensor(index, index_1)
+            return (mul_1,)
+
+        x = torch.zeros(2, 209985).to(torch.int64)
+        opt_fn = torch._dynamo.optimize("inductor")(fn)
+        _, code = run_and_get_cpp_code(opt_fn, x)
+        FileCheck().check_count(
+            "return at::vec::VectorizedN<int64_t,2>::loadu(tmpbuf.data(),",
+            2,
+            exactly=True,
+        ).run(code)
+
 
 if __name__ == "__main__":
     from torch._inductor.test_case import run_tests

torch/_inductor/codegen/cpp.py

@@ -2336,7 +2336,7 @@ def vec_to_array(vec_var: CppCSEVariable) -> CppCSEVariable:
             assert vec_var.is_vec
             code = BracesBuffer()
             code.writeline("[&]")
-            with self.swap_buffers(code), code.indent():
+            with code.indent():
                 vec_dtype = vec_var.dtype
                 assert vec_dtype is not None
                 if vec_dtype == torch.bool:
@@ -2357,7 +2357,7 @@ def vec_to_array(vec_var: CppCSEVariable) -> CppCSEVariable:
         assert opt_ctx is not None
         code = BracesBuffer()
         code.writeline("[&]")
-        with self.swap_buffers(code), code.indent():
+        with code.indent():
             result_size = get_result_size(dtype)
             result_declare = (
                 f"__at_align__ std::array<{DTYPE_TO_CPP[dtype]}, {result_size}> tmpbuf;"