https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104377
--- Comment #4 from Feng Xue <fxue at os dot amperecomputing.com> --- (In reply to Martin Jambor from comment #2) > (In reply to Feng Xue from comment #1) > > > > OK. I does missed something. Here we could not hold assumption that > > ipcp_decision_stage() only sees raw cgraph node, since sometime in the > > future some new ipa pass may be added prior to ipa-cp, and this pass > > introduces clone node. > > Right, initially IPA-SRA was developed as a pass before IPA-CP and it > may well be that we decide to swap the order again. > > > > > However, there is a questionable point about the code snippet > > > > if (!node->can_change_signature > > || old_adj->op != IPA_PARAM_OP_COPY > > || (!known_csts[old_adj->base_index] > > && ipa_is_param_used (info, old_adj->base_index))) > > > > In ipa-cp, known_csts is for the node, has no relation to the node's origin > > node, but here it is accessed via index of the latter (old_adj->base_index), > > will this cause out-of-bound error? > > I think the code is correct. Assume IPA-SRA running before IPA-CP, and > we're compiling a function with two argument, with indices 0 and 1. > > Analysis phases of both passes run before the IPA (WPA) phases of > either. This is important to keep in mind. > > IPA SRA removes the first one with index zero as useless, IPA-CP wants > to remove the second one with index 1, possibly because it is constant > everywhere. In oder to that it has to combine the pre-existing > adjustments with its own changes. > > Before create_specialized_node, the pass checks whether previous > passes did not kill some parameters and stops caring about them, but > it does not re-index anything, all lattices, jump functions, > everything, still keep their positions (and thus indices) they got in > the analysis phase. > > Then create_specialized_node hits this loop. For i=0 encounters an > old_adj element that actually describes the parameter which originally > had index 1. The pass looks up in base_index what the original > (pre-IPA) index of the parameter was (1) and uses those to look up in > its own structures whether it wants to remove it or not. > > Bounds should be always OK, base_index should never be greater than > the original pre-IPA number of parameters (-1) and known_csts should > always have just as many parameters. > > Does that make sense? Yes. Thanks for your explanation.
