On Mon, Mar 26, 2012 at 9:02 PM, Richard Sandiford
>> * cse.c (cse_canonicalized_basic_blocks): New simple bitmap to
>> tag basic blocks that have already been traversed at least once,
>> so that all insns have been canonicalized.
>> (cse_insn): Call canonicalize_insn only if the basic block that
>> contains insn is visited for the first time.
>> (cse_extended_basic_block): After visiting all insns in a basic
>> block, mark the block in cse_canonicalized_basic_blocks.
>> (cse_main): Setup and destroy cse_canonicalized_basic_blocks.
> OK, thanks (without the microoptimisation, as you say).
Thanks, will commit!
> Out of curiosity, do you still see this bit as useful:
> /* We potentially will process this insn many times. Therefore,
> drop the REG_EQUAL note if it is equal to the SET_SRC of the
> unique set in INSN.
> Do not do so if the REG_EQUAL note is for a STRICT_LOW_PART,
> because cse_insn handles those specially. */
> ? Does "many times" mean in CSE, or later?
It means "in CSE", yes. But with the patch to canonicalize only once,
I suppose this can go away again.
Having said that, I do believe it would be good if we avoid having
REG_EQUAL notes that are rtx_equal_p to the SET_SRC. This happens
quite frequently after CSE. I'm not sure how to clean them up.