On May 14, 2018 2:04:38 AM PDT, David Laight <david.lai...@aculab.com> wrote: >From: H. Peter Anvin >> Sent: 11 May 2018 19:54 >> >> On 05/10/18 09:39, David Laight wrote: >> > From: Alexey Dobriyan >> >> Sent: 07 May 2018 22:38 >> >> >> >> Use INT3 instead of NOP. All that padding between functions is >> >> an illegal area, no legitimate code should jump into it. >> >> >> >> I've checked x86_64 allyesconfig disassembly, all changes looks >sane: >> >> INT3 is only used after RET or unconditional JMP. >> > >> > I thought there was a performance penalty (on at least some cpu) >> > depending on the number of and the actual instructions used for >padding. >> > >> > I believe that is why gcc generates a small number of very long >'nop' >> > instructions when padding code. >> > >> >> There is a performance penalty for using NOP instructions *in the >> fallthrough case.* In the case where the padding is never supposed >to >> be executed, which is what we're talking about here, it is >irrelevant. > >Not completely irrelevant, the instructions stream gets fetched and >decoded >beyond the jmp/ret at the end of the function. >At some point the cpu will decode the jmp/ret and fetch/decode from the >target address. >I guess it won't try to speculatively execute the 'pad' instructions >- but you can never really tell! > > David
The CPU doesn't speculate down past an unconditional control transfer. Doing so would be idiotic. -- Sent from my Android device with K-9 Mail. Please excuse my brevity.
