Packet-oriented data transmission is gaining more and more importance in wireless communication systems. Typically, data transmission is not strictly delay-sensitive but requires a virtually error-free link. In order to provide such level of reliability over wireless channels, affected by propagation impairments such as fading, Automatic Retransmission reQuest (ARQ) schemes can be combined with channel coding (Hybrid ARQ). In brief, when fading varies slowly over the duration of a codeword, coding takes care of the channel noise while retransmissions take care of bad channel conditions (deep fades). In this work we study the throughput achievable by H-ARQ schemes based on incremental redundancy over a block-fading channel. We provide an information-theoretic analysis assuming binary random coding and typical-set decoding. Then, we study the performance of Low-Density Parity-Check code ensembles with iterative belief-propagation decoding and show that, assuming infinite block length, LDPC codes yield almost optimal performance. Unfortunately, practical finite-length LDPC codes incur a considerable performance loss with respect to their infinite-length counterpart. In order to reduce this performance loss, we propose two very effective methods: 1) using special LDPC ensembles designed to provide good frame-error rate (rather than just good iterative decoding threshold); 2) using an outer selective-repeat protocol acting on smaller packets of information bits. Surprisingly, these two apparently very different methods yield almost the same performance gain and recover a considerable fraction of the optimal throughput, thus making practical finite-length LDPC codes very attractive for data wireless communications based on incremental redundancy H-ARQ schemes.