GN: Use implicit dependency for binary input deps

tools/gn/ninja_binary_target_writer_unittest.cc

Index: tools/gn/ninja_binary_target_writer_unittest.cc
diff --git a/tools/gn/ninja_binary_target_writer_unittest.cc b/tools/gn/ninja_binary_target_writer_unittest.cc
index f2297fbf206e22f3af71e2a3e119a2024004acac..eadd2f769197349530236a1ecce14e7048823403 100644
--- a/tools/gn/ninja_binary_target_writer_unittest.cc
+++ b/tools/gn/ninja_binary_target_writer_unittest.cc
@@ -294,9 +294,9 @@ TEST(NinjaBinaryTargetWriter, OutputExtensionAndInputDeps) {
       "target_output_name = libshlib\n"
       "\n"
       "build obj/foo/libshlib.input1.o: cxx ../../foo/input1.cc"
-        " || obj/foo/action.stamp\n"
+        " | obj/foo/action.stamp\n"

[brettw, 2016/06/16 19:41:34]

Actually, we specifically don't want to change this case. This is what
order-only deps were designed for.

What I was imagining when I talked about this is how the "inputs" list in a
binary target is treated. Currently they also appear in the order only deps, but
I think we want to split them out to be implicit deps. This should apply only to
the explicit inputs on a target, and not the transitive dependencies of it. So
it would apply to these things:

source_set("foo") {
  sources = [ "foo", "bar" ]
  inputs = [ "baz.dat" ]
}

Changing all actions to be implicit dependencies of binary targets that depend
on them in some way (what it looks like you're doing) will reduce build
parallelism significantly.

[Petr Hosek, 2016/06/16 23:11:47]

Acknowledged.

       "build obj/foo/libshlib.input2.o: cxx ../../foo/input2.cc"
-        " || obj/foo/action.stamp\n"
+        " | obj/foo/action.stamp\n"
       "\n"
       "build ./libshlib.so.6: solink obj/foo/libshlib.input1.o "
       // The order-only dependency here is stricly unnecessary since the