On Wed, Jul 13, 2011 at 11:49 PM, Steven Bosscher <stevenb....@gmail.com> wrote: > On Wed, Jul 13, 2011 at 4:34 PM, Richard Guenther > <richard.guent...@gmail.com> wrote: >> On Wed, Jul 13, 2011 at 3:13 PM, Andreas Krebbel >> <kreb...@linux.vnet.ibm.com> wrote: >>> Hi, >>> >>> the widening_mul pass might increase the number of multiplications in >>> the code by transforming >>> >>> a = b * c >>> d = a + 2 >>> e = a + 3 >>> >>> into: >>> >>> d = b * c + 2 >>> e = b * c + 3 >>> >>> under the assumption that an FMA instruction is not more expensive >>> than a simple add. This certainly isn't always true. While e.g. on >>> s390 an fma is indeed not slower than an add execution-wise it has >>> disadvantages regarding instruction grouping. It doesn't group with >>> any other instruction what has a major impact on the instruction >>> dispatch bandwidth. >>> >>> The following patch tries to figure out the costs for adds, mults and >>> fmas by building an RTX and asking the backends cost function in order >>> to estimate whether it is whorthwhile doing the transformation. >>> >>> With that patch the 436.cactus hotloop contains 28 less >>> multiplications than before increasing performance slightly (~2%). >>> >>> Bootstrapped and regtested on x86_64 and s390x. >> >> Ick ;) > > +1 > >> Maybe this is finally the time to introduce target hook(s) to >> get us back costs for trees? For this case we'd need two >> actually, or just one - dependent on what finegrained information >> we pass. Choices: >> >> tree_code_cost (enum tree_code) >> tree_code_cost (enum tree_code, enum machine_mode mode) >> unary_cost (enum tree_code, tree actual_arg0) // args will be mostly >> SSA names or constants, but at least they are typed - works for >> mixed-typed operations >> binary_cost (...) >> ... >> unary_cost (enum tree_code, enum tree_code arg0_kind) // constant >> vs. non-constant arg, but lacks type/mode > > Or maybe add a cost function for all named insns (i.e. > http://gcc.gnu.org/onlinedocs/gccint/Standard-Names.html#Standard-Names)? > I think that any form of lower GIMPLE will not be so low level that > more combinations will exist than the available named patterns. It > should be possible to write a gen* tool using rtx_costs to compute > some useful cost metric for all named patterns. How complicated that > could be (modes, reg vs. mem, etc.), I don't know... But at least that > way we don't end up with multiple target costs depending on the IR in > use.
Yeah, it occured to me as well that when we look for supportable operations via optabs the same mechanism should also be able to provide a cost, maybe as simple as attaching one to the named expanders. Generating RTL from GIMPLE passes just to be able to use rtx_cost is, well ... gross. Yes, we do it in IVOPTs (and that case is even more complex), but I don't think we want to start doing it elsewhere (look how the vectorizer for example uses new target hooks instead of generating vectorized RTL and then using rtx_cost). Richard. > Ciao! > Steven >
