It is alright to apply read trimming thing only for those samples whose reads have low quality bases in the end. Samples with different read lengths can be used for expression analysis as in the end you will be using RPKM normalised values that will normalize the read counts for genes by dividing the number of reads mapped by the length of the coding region and number of total reads mapped for that sample. This way even if trimming the low quality reads have narrowed down the number of reads mapped (it may happen that trimming made the read really small so that it can't be confidently aligned against the genome and discarded. Though it will rarely happen as you have long read Hiseq PE data) , the RPKM will take care of it.