On Sun, Aug 6, 2017 at 7:44 PM, Luc Van Oostenryck <luc.vanoostenryck@xxxxxxxxx> wrote: >> >> May I ask the question that, the SSA form you are building >> from this codes. Does it violate the SSA dominance property > > Of course not. Unless there is a bug, that is. I am very glad to heard that :-) > >> In other word, does it produce the "usage before >> define" IR similar to this? >> >> and.32 %r3 <- %r4, $-65 >> or.32 %r4 <- %r3, $64 > > It gives the code as I showed as reply to Dibyendu's mail. > For example, his code gives: > foo: > .L0: > <entry-point> > and.32 %r2 <- UNDEF, $-65 > or.32 %r3 <- %r2, $64 > lsr.32 %r4 <- %r3, $6 > cast.1 %r5 <- (32) %r4 > cast.32 %r6 <- (1) %r5 > setne.32 %r7 <- %r6, $1 > ret.32 %r7 > > with UNDEF being a new type of pseudo. OK. So far so good. > > On purpose, nothing is done yet with the UNDEFs. > >> That I am very curios how it was done. See the above >> question. > > IMO, the best is to read the article. My code is > still very close to the article's pseudo-code. > > Basically, it's a lazy approach. Things are done > when possible and are delayed if not. So you are building the SSA for the memory variable I assume? Directly from the linearize process? > More specifically, you can do what is needed in a BB > once you have all the parents, otherwise you need to > delay things. The article call this 'BB is sealed' > and there is a 'seal' operation that need to be called > to process what was delayed. > Complications arise with gotos (and the article doesn't > talk about them) but I've added a solution for it > (but I would like to return to it later). I am missing the obvious. Where is this article? > > From what I've seen/heard this algo have been pretty well > received. Personnaly, I'm very happy with its simplicity > and its result. Great! > > One thing that is missing is the more complex cases > with trivial phi-nodes. Can you clarify? > > Another thing I would like to change is the ptrmap > implementation I've added: it need to be hash-table based > (but it's only an implementation detail that matters > for performance on really big functions). Which ptrmap? The one used for CSE? Chris -- To unsubscribe from this list: send the line "unsubscribe linux-sparse" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
