This is the mail archive of the mailing 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]

Re: [PATCH] Improve lim for pulling loads from global variable out of loops


> > > > 	* tree-ssa-loop-im.c (stmt_cost): Gimple variables should be
> > > > 	pulled out too as they cause memory accesses.
> > > > 
> > > Your ChangeLog entry doesn't match the code.  Also, testing only for
> > > is_gimple_addressable will not include global variables.  You want to
> > > test is_global_var.
> > 
> > Diego, I think you are wrong in both points -- is_gimple_addressable
> > includes global variables (any variables, in fact), and the changelog
> > entry seems to match the change.
> > 
> You're right, I didn't express myself correctly.  Andrew said he wanted
> to consider global variables expensive.  Globals are addressable, but
> not every addressable is a global.
> > However the patch is wrong -- is_gimple_addressable also matches SSA
> > names.  So the right patch should be
> > 
> I thought the intent was to only mark globals as expensive?  I think I
> would just say:

the intention of the code is "everything that performs load from
memory".  I.e. addressable && ! SSA name.

> if (!is_gimple_reg (rhs))
>    cost += 20;

This would cause us to count any operation except for simple reg = reg
copy as expensive, which definitely is not what is wanted.


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