Steven Bosscher - Re: [dataflow] [RFC] Remove many almost useless DCE passes (original) (raw)

This is the mail archive of the gcc-patches@gcc.gnu.orgmailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

On 2/27/07, Vladimir Makarov vmakarov@redhat.com wrote:

Good because SPECFP2000 for ppc64 is 0.5% better (SPECInt2000 is the same) on the branch than on the mainline at the last merge point. And this patch could remove this achievement.

Or it might not make a difference at all.

Which is actually more likely IMHO, but I'll let the numbers speak for me, instead of doing some hand waiving and speculating.

You seem to believe that mainline as-is with all its delete_trivially_dead_insns calls makes sense.

It really doesn't.

Bonzini and I have gathered the numbers to show this. It makes almost no sense on the trunk and no sense at all on the dataflow branch.

On the trunk, removing trivially dead code often is helpful because delete_trivially_dead_insns does not clean up dead code very well.

On the dataflow branch, only the delete_trivially_dead_insns calls after jump, fwprop1, gcse1, and cse2 remove dead code. All other d_t_d_i calls do nothing but waste time.

The reason for this, is that the DCE pass on the dataflow branch is much brighter than delete_trivially_dead_insns (and often smarter than what is in flow.c on the trunk, although it apparently still needs improvements in some areas, too).

Gr. Steven

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]