On Tue, Jul 2, 2019 at 3:07 PM Peter Zijlstra <pet...@infradead.org> wrote: > > > Hi Nick, > > That is good news; and I'll strive to read the email in more detail > in the morning when there is a better chance of me actually > understanding some of it :-) > > But his here is something I felt needed clarification: > > On Tue, Jul 02, 2019 at 01:53:51PM -0700, Nick Desaulniers wrote: > > Of interest are the disassembled __jump_table entries; in groups of > > three, there is a group for which the second element is duplicated > > with a previous group. This is bad because (as explained by Peter in > > https://lkml.org/lkml/2019/6/27/118) the triples are in the form (code > > location, jump target, pointer to key). Duplicate or repeated jump > > targets are unexpected, and will lead to incorrect control flow after > > patching such code locations. > > > Also, the jump target should be 0x7 bytes ahead of the location, IIUC. > > Even if you mean 'at least' I'm fairly sure this is not correct. The > instruction at the 'code location' is either a jmp.d32 or a nop5 (both 5 > bytes). The target must (obviously) be at an instruction boundary, but > really can be anywhere (it is compiler generated after all). >
Got it. Issue should be fixed outright with: https://reviews.llvm.org/D64101 (I figured out how to fix the loop unroller, so now we can proceed with the optimization instead of the base conservative case). -- Thanks, ~Nick Desaulniers
