// same way, and each type of Resource must be fully parsed or skipped before
// proceeding to the next type of Resource.
//
// Parser is safe to copy to preserve the parsing state.
//
// Note that there is no requirement to fully skip or parse the message.
type Parser struct {
	msg    []byte
	header header

	section         section
	off             int
	index           int
	resHeaderValid  bool
	resHeaderOffset int

//
//
// A parameterized test fixture must be derived from testing::Test and from
// testing::WithParamInterface<T>, where T is the type of the parameter
// values. Inheriting from TestWithParam<T> satisfies that requirement because
// TestWithParam<T> inherits from both Test and WithParamInterface. In more
// complicated hierarchies, however, it is occasionally useful to inherit
// separately from Test and WithParamInterface. For example:

//
//
// A parameterized test fixture must be derived from testing::Test and from
// testing::WithParamInterface<T>, where T is the type of the parameter
// values. Inheriting from TestWithParam<T> satisfies that requirement because
// TestWithParam<T> inherits from both Test and WithParamInterface. In more
// complicated hierarchies, however, it is occasionally useful to inherit
// separately from Test and WithParamInterface. For example:

					// ADDQconst is added here because we want to treat it as resultInArg0 for
					// the purposes of desired registers, even though it is not an absolute requirement.
					// This is because we'd rather implement it as ADDQ instead of LEAQ.
					// Same for ADDLconst
					// Select0 is added here to propagate the desired register to the tuple-generating instruction.

to such variables for each replica are redundant and can be replaced by writing
a single value from first replica.

The benefit of this optimization is reduced memory requirement on host. For
multiple writes (one from each replica) to such variables, the host would
allocate buffer space to receive the device output from all replicas, which is