------- Additional Comments From mark at codesourcery dot com 2005-03-03 19:34 ------- Subject: Re: [3.4 Regression] const/pure functions result in bad asm
ebotcazou at gcc dot gnu dot org wrote: > ------- Additional Comments From ebotcazou at gcc dot gnu dot org 2005-03-03 > 19:25 ------- > >>It really seems like the C++ front end is doing the right thing, >>abstractly -- these functions don't have side-effects! So, either the >>inliner or stabilize reference seems like it needs fixing. Maybe the >>latter should recognize CALL_EPRs? > > > I agree that the C++ FE is theoritically right, but other FEs appear to > already > have realized that the implementation constraints trumpet the theory here: on > mainline, the C, Java and Fortran 95 FEs set TREE_SIDE_EFFECTS on CALL_EXPRs. Interesting. So, we have some checks in gcc/tree.c to try to avoid setting TREE_SIDE_EFFECTS, and then the FEs all set it anyhow. :-( > Special-casing CALL_EXPRs in stabilize_reference could indeed be the > solution. > We could even be clever and check whether tree inlining is enabled and, if so, > whether it has already occurred. Does that sound plausible? I don't think I'd try to be that clever. We might want the stabilization to occur even in other cases. In looking at it more closely, it definitely looks like stabilize_reference should deal with CALL_EXPRs. (I guess I was thinking that the problem with the duplicate labels could be avoided in the inliner by generating fresh labels. But, maybe that's not going to work for other reasons.) >>If it's not possible to fix this, then what we should do is modify >>build3 to mark CALL_EXPRs as having side-effects. Right now, it makrs >>them based on the flags for the function, so this problem isn't really >>C++-specific; it probably effects all languages except those whose front >>ends set TREE_SIDE_EFFECTS additionally themselves. (Unless C++ is >>changing out the operands to the CALL_EXPR after its created.) > > Right. OK, that's the fallback. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17972
