> On Fri, Dec 12, 2008 at 12:29 AM, Martin Jambor <mjam...@suse.cz> wrote: > > Hi, > > > > today I have encountered an unpleasant problem with the function > > get_ref_base_and_extent() when it claimed a known and constant offset > > for the expression insn_4(D)->u.fld[arg.82_3].rt_rtvec. (arg being a > > default_def parameter of the function, insn is an rtx). Moreover, it > > also returned constant size and max_size, all three equal to 64 > > (bits). > > > > This is certainly wrong (I believe the function got confused by > > unions) but after looking at the function and what it did in debugger, > > I grew unsure what the expected behavior is. The two alternatives I > > consider possibly correct are returned offset equal to -1 or, > > alternatively, offset equal to the offset of the array (+ offset of > > rt_rtvec which is zero) and max_size equal either to the size of the > > array or -1 if it is unknown. > > > > At the moment, the function never returns offset equal to -1, the > > comment above it does not even mention such possibility and, from the > > limited research I did, its callers do not expect and check for it. > > However, at the same time, the special handling of non-constants in > > array indices after the label "done" does not trigger in this > > particular case (here I suspect the unions come to play because it is > > the last part of the conjunction that evaluates to false). > > > > Which (or what else) is the correct semantics of the function? Both > > make sense to me (I would prefer the former but I suspect other users > > rely on the latter). What would be the correct fix, when a union > > field is itself a record ending with a variable-length array? How > > much would I pessimize things if I just returned -1 max_size if both a > > non-constant index and a union was encountered? > > > > Or did I miss something else? > > The function is supposed to return -1 for max_size in case the access > accesses a variable array index of an array of unknown size. > > insn_4(D)->u.fld[arg.82_3].rt_rtvec accesses struct rtvec_def -- in which > case the access _does_ have known size (an int plus one rtx pointer). > But maybe I am missing context here?
Hmm, isn't this the usual problem with arrays running past the end of sturcture? I guess we need to special case these for needs of SRA as you can't easilly know size of the array even if it is explicitely written in the program. We similarly special case those in VRP, right? Jan > > Richard. > > > Thanks a lot in advance, > > > > Martin > >
