// If the op is the pass-through op with (3x) smaller output, the dequantize
    // op can be pushed down to the single result of this op.
    if (!llvm::dyn_cast<mlir::SameScalesOpInterface>(passthrough_op) ||
        passthrough_op->getNumResults() != 1) {
      return failure();
    }
    // Only push down the dequantize op when the output is smaller, so that it
    // can have smaller memory usage.
    auto input_type =

    //     imagine a line separating the first block from the second, we can proceed by exchanging
    //     the smaller of these blocks with the far end of the other one. That leaves us with a
    //     smaller version of the same problem.
    //     Say we are rotating abcdefgh by 5. We start with abcde|fgh. The smaller block is [fgh]:
    //     [abc]de|[fgh] -> [fgh]de|[abc]. Now [fgh] is in the right place, but we need to swap [de]

  auto id_filtered2 =
      StringAttr::get(&context, "dir/tensorflow/python/filtered_file_B.py");
  auto loc_filtered2 = FileLineColLoc::get(&context, id_filtered2, 0, 0);

  // Build a small stack for each error; the MLIR diagnostic filtering will
  // surface a location that would otherwise be filtered if it is the only
  // location associated with an error; therefore we need a combinatination of

    // This may have performance implications, however the assumption is that class generators are global scoped objects that are created once and in very small numbers
    private static final Map<String, Throwable> SUFFIXES = new ConcurrentHashMap<>();

    /**
     * Registers the given suffix as in use for generated class names.
     */

func addUniqueMap[T comparable](typ *abi.Type) *uniqueMap[T] {
	// Create a map for T and try to register it. We could
	// race with someone else, but that's fine; it's one
	// small, stray allocation. The number of allocations
	// this can create is bounded by a small constant.
	m := &uniqueMap[T]{
		HashTrieMap: concurrent.NewHashTrieMap[T, weak.Pointer[T]](),
		cloneSeq:    makeCloneSeq(typ),
	}

    //     imagine a line separating the first block from the second, we can proceed by exchanging
    //     the smaller of these blocks with the far end of the other one. That leaves us with a
    //     smaller version of the same problem.
    //     Say we are rotating abcdefgh by 5. We start with abcde|fgh. The smaller block is [fgh]:
    //     [abc]de|[fgh] -> [fgh]de|[abc]. Now [fgh] is in the right place, but we need to swap [de]

				// confusing the benchmarking framework for small b.N.
				b.StopTimer()
				stop()
			})
		}
	}

	// Measure the cost of counting goroutines
	b.Run("small-nil", run(func() bool {
		GoroutineProfile(nil)
		return true
	}))

	// Measure the cost with a small set of goroutines
	n := NumGoroutine()
	p := make([]StackRecord, 2*n+2*GOMAXPROCS(0))
	b.Run("small", run(func() bool {

	}{
		// dual-stack:
		{"dual IPv4 only, but mask too small. Default node-mask", "10.0.0.16/29", nil, false},
		{"dual IPv4 only, but mask too small. Configured node-mask", "10.0.0.16/24", []kubeadmapi.Arg{{Name: "node-cidr-mask-size-ipv4", Value: "23"}}, false},
		{"dual IPv6 only, but mask too small. Default node-mask", "2001:db8::1/112", nil, false},

build:windows --cxxopt=/std:c++17
build:windows --host_cxxopt=/std:c++17

# On windows, we still link everything into a single DLL.
build:windows --config=monolithic

# On linux, we dynamically link small amount of kernels
build:linux --config=dynamic_kernels

# Make sure to include as little of windows.h as possible
build:windows --copt=-DWIN32_LEAN_AND_MEAN
build:windows --host_copt=-DWIN32_LEAN_AND_MEAN

        "//tensorflow/core:test",
        "//tensorflow/core:test_main",
        "//tensorflow/core:testlib",
    ],
)

tf_cc_test(
    name = "framework_while_gradients_test",
    size = "small",
    srcs = ["framework/while_gradients_test.cc"],
    deps = [
        ":cc_ops",
        ":client_session",
        ":grad_op_registry",
        ":grad_ops",
        ":gradients",
        ":testutil",