On Mon, Aug 6, 2012 at 1:07 PM, Paolo Bonzini <bonz...@gnu.org> wrote:
> Il 06/08/2012 08:54, Steven Bosscher ha scritto:
>> Hello,
>>
>> In PR54146, ifcvt spends a lot of time just clearing memory. This
>> patch changes the value maps to pointer-maps to fix this issue.
>>
>> Bootstrapped&tested on x86_64-unknown-linux-gnu. OK?
>
> Nice, but perhaps we need a sparsemap to do even better?
If you mean sparseset: no, it won't do any better.
My first idea was to use a sparseset, but:
1. The amount of memory allocated is simply too big. In fact, it'd be
even worse than before: 4*max_regno*sizeof(rtx) instead of "only"
2*max_regno*sizeof(rtx).
2. sparseset has the same problem of memory clearing (for valgrind,
see sparseset_alloc).
Remember, we can call this function cond_move_convert_if_block
O(n_basic_block) times: On a CFG with mostly regular control flow,
i.e. no or few computed jumps, jump tables, and exception handlers,
the number of IF-THEN-{,ELSE_}-JOIN blocks will be O(n_basic_blocks)
and cond_move_convert_if_block is called on almost all of them.
What could work, is to allocate a sparseset once at the beginning of
if_convert, but I don't see any good reason to add a pair of global
variables for this. My patch reduces the time spent in ifcvt to almost
nothing (down from ~250s, which is still not so bad, considering we're
looking at a function with >260000 basic blocks!).
Ciao!
Steven
