2017-03-15 13:15 GMT+03:00 Martin Liška <mli...@suse.cz>: > On 03/14/2017 11:27 PM, Ilya Enkovich wrote: >> >> 2017-03-10 16:15 GMT+03:00 Martin Liška <mli...@suse.cz>: >>> >>> Hello. >>> >>> Currently, __builtin_ia32_bndret is used for all functions that have >>> non-void >>> return type. I think the right fix is to return bounds just for a bounded >>> type. >>> >>> Patch can bootstrap on x86_64-linux-gnu and survives regression tests. >>> >>> Ready to be installed? >>> Martin >> >> >> Hi, >> >> The fix makes sense and allows you to pass the test but it seems like we >> still have the issue in inlining which can result in bndret call with >> wrong arg. > > > Hi. > > The test I added does probably what you mean: a return value of integer is > casted > to a pointer type. Or?
I suspect problem may still exist when you cast function, not just returned value. Inline is supposed to always remove retbnd corresponding to expanded call. You avoid the problem in the testcase by not producing retbnd call. But in case of calling a function casted to another function type we might still hit this issue in inline. I'll try to make a test. Thanks, Ilya > >> Do you avoid all such cases by this fix? Can we still have similar problem >> if cast function with integer return type to a function with pointer >> return type >> and the inline the call (not sure if we can inline such calls)? >> >> I think this fix still should go to trunk anyway. > > > Good, may I consider this as patch approval? > Thanks, > Martin > >> >> Thanks, >> Ilya >> >
