https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65368
--- Comment #7 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Jakub Jelinek from comment #6) > To some extent BZHI is a special case of BEXTR, but I'm afraid any > generalization is much harder due to the weirdo encoding of the operands, > unless both the START and LEN are CONST_INTs (and if they are constants, it > is questionable if BEXTR or BZHI are the best instructions to use, because > both take just register for the LEN or START/LEN pair, so one would need to > set some register to a constant and then perform BEXTR/BZHI, so two > instructions, while there is always the option to perform AND and right > shift (or right shift and AND). Probably I didn't form the question in the right way - I was trying to point out, if we also describe BEXTR without unspec (in a similar way as your proposed patch describes BZHI), then combine will be able to propagate eventual immediates to the insn, and in this case reduce BEXTR with START=0 and LEN=n to an AND.
