Avoid redundant calculation of hash data during join probe stage

[windtalker]

Enhancement

In join probe stage, for each input block, it will be wrapped inside ProbeProcessInfo, this is because if the build side has many duplicated key entries, the probe stage may expand the input block greatly, and inorder to keep the max-row-size of a output block, ProbeProcessInfo is used to support probe only part of the input block. That is to say for an input block, it may be probed multiple times, each time of part of the data in the block is processed.

For each probe, it will call probeBlockImplTypeCase to do the probe, and inside that function, it will calculate the hash data of the whole block:

tiflash/dbms/src/Interpreters/JoinPartition.cpp

Lines 1480 to 1494 in e597b78

	if (join_build_info.needVirtualDispatchForProbeBlock())
	{
	assert(!(join_build_info.restore_round > 0 && join_build_info.enable_fine_grained_shuffle));
	/// TODO: consider adding a virtual column in Sender side to avoid computing cost and potential inconsistency by heterogeneous envs(AMD64, ARM64)
	/// Note: 1. Not sure, if inconsistency will do happen in heterogeneous envs
	/// 2. Virtual column would take up a little more network bandwidth, might lead to poor performance if network was bottleneck
	/// Currently, the computation cost is tolerable, since it's a very simple crc32 hash algorithm, and heterogeneous envs support is not considered
	computeDispatchHash(
	rows,
	key_columns,
	collators,
	sort_key_containers,
	join_build_info.restore_round,
	build_hash);
	}

If a block is probed multiple times, the hash data will be calculated multiple times, which is meaningless and redundant.