Message ID | 20190409212517.7321-1-linux@rasmusvillemoes.dk (mailing list archive) |
---|---|
Headers | show |
Series | implement DYNAMIC_DEBUG_RELATIVE_POINTERS | expand |
On 09/04/2019 23.25, Rasmus Villemoes wrote: > While refreshing these patches, which were orignally just targeted at > x86-64, it occured to me that despite the implementation relying on > inline asm, there's nothing x86 specific about it, and indeed it seems > to work out-of-the-box for ppc64 and arm64 as well, but those have > only been compile-tested. So, apart from the Clang build failures for non-x86, I now also got a report that gcc 4.8 miscompiles this stuff in some cases [1], even for x86 - gcc 4.9 does not seem to have the problem. So, given that the 5.2 merge window just opened, I suppose this is the point where I should pull the plug on this experiment :( Rasmus [1] Specifically, the problem manifested in net/ipv4/tcp_input.c: Both uses of the static inline inet_csk_clear_xmit_timer() pass a compile-time constant 'what', so the ifs get folded away and both uses are completely inlined. Yet, gcc still decides to emit a copy of the final 'else' branch of inet_csk_clear_xmit_timer() as its own inet_csk_reset_xmit_timer.part.55 function, which is of course unused. And despite the asm() that defines the ddebug descriptor being an "asm volatile", gcc thinks it's fine to elide that (the code path is unreachable, after all....), so the entire asm for that function is .section .text.unlikely .type inet_csk_reset_xmit_timer.part.55, @function inet_csk_reset_xmit_timer.part.55: movq $.LC1, %rsi #, movq $__UNIQUE_ID_ddebug160, %rdi #, xorl %eax, %eax # jmp __dynamic_pr_debug # .size inet_csk_reset_xmit_timer.part.55, .-inet_csk_reset_xmit_timer.part.55 which of course fails to link since the symbol __UNIQUE_ID_ddebug160 is nowhere defined.
* Rasmus Villemoes <linux@rasmusvillemoes.dk> wrote: > On 09/04/2019 23.25, Rasmus Villemoes wrote: > > > While refreshing these patches, which were orignally just targeted at > > x86-64, it occured to me that despite the implementation relying on > > inline asm, there's nothing x86 specific about it, and indeed it seems > > to work out-of-the-box for ppc64 and arm64 as well, but those have > > only been compile-tested. > > So, apart from the Clang build failures for non-x86, I now also got a > report that gcc 4.8 miscompiles this stuff in some cases [1], even for > x86 - gcc 4.9 does not seem to have the problem. So, given that the 5.2 > merge window just opened, I suppose this is the point where I should > pull the plug on this experiment :( > > Rasmus > > [1] Specifically, the problem manifested in net/ipv4/tcp_input.c: Both > uses of the static inline inet_csk_clear_xmit_timer() pass a > compile-time constant 'what', so the ifs get folded away and both uses > are completely inlined. Yet, gcc still decides to emit a copy of the > final 'else' branch of inet_csk_clear_xmit_timer() as its own > inet_csk_reset_xmit_timer.part.55 function, which is of course unused. > And despite the asm() that defines the ddebug descriptor being an "asm > volatile", gcc thinks it's fine to elide that (the code path is > unreachable, after all....), so the entire asm for that function is > > .section .text.unlikely > .type inet_csk_reset_xmit_timer.part.55, @function > inet_csk_reset_xmit_timer.part.55: > movq $.LC1, %rsi #, > movq $__UNIQUE_ID_ddebug160, %rdi #, > xorl %eax, %eax # > jmp __dynamic_pr_debug # > .size inet_csk_reset_xmit_timer.part.55, > .-inet_csk_reset_xmit_timer.part.55 > > which of course fails to link since the symbol __UNIQUE_ID_ddebug160 is > nowhere defined. It's sad to see such nice data footprint savings go the way of the dodo just because GCC 4.8 is buggy. The current compatibility cut-off is GCC 4.6: GNU C 4.6 gcc --version Do we know where the GCC bug was fixed, was it in GCC 4.9? According to the GCC release dates: https://www.gnu.org/software/gcc/releases.html 4.8.0 was released in early-2013, while 4.9.0 was released in early-2014. So the tooling compatibility window for latest upstream would narrow from ~6 years to ~5 years. Thanks, Ingo
On 06/05/2019 09.05, Ingo Molnar wrote: > > > It's sad to see such nice data footprint savings go the way of the dodo > just because GCC 4.8 is buggy. > > The current compatibility cut-off is GCC 4.6: > > GNU C 4.6 gcc --version > > Do we know where the GCC bug was fixed, was it in GCC 4.9? Not sure. The report was from a build on CentOS with gcc 4.8.5, so I tried installing the gcc-4.8 package on my Ubuntu machine and could reproduce. Then I tried installed gcc-4.9, and after disabling CONFIG_RETPOLINE (both CentOS and Ubuntu carry backported retpoline support in their 4.8, but apparently not 4.9), I could see that the problem was gone. But whether it's gone because it no longer elides an asm volatile() on a code path it otherwise emits code for, or because it simply doesn't emit the unused static inline() at all I don't know. I thought 0day also tested a range of supported compiler versions, so I was rather surprised at getting this report at all. But I suppose the arch/config matrix is already pretty huge. Anyway, the bug certainly doesn't exist in any of the gcc versions 0day does test. I _am_ bending the C rules a bit with the "extern some_var; asm volatile(".section some_section\nsome_var: blabla");". I should probably ask on the gcc list whether this way of defining a local symbol in inline assembly and referring to it from C is supposed to work, or it just happens to work by chance. Rasmus
* Rasmus Villemoes <linux@rasmusvillemoes.dk> wrote: > I _am_ bending the C rules a bit with the "extern some_var; asm > volatile(".section some_section\nsome_var: blabla");". I should > probably ask on the gcc list whether this way of defining a local > symbol in inline assembly and referring to it from C is supposed to > work, or it just happens to work by chance. Doing that would be rather useful I think. Thanks, Ingo
On Mon, May 06, 2019 at 09:34:55AM +0200, Rasmus Villemoes wrote: > I _am_ bending the C rules a bit with the "extern some_var; asm > volatile(".section some_section\nsome_var: blabla");". I should probably > ask on the gcc list whether this way of defining a local symbol in > inline assembly and referring to it from C is supposed to work, or it > just happens to work by chance. It only works by chance. There is no way GCC can know the asm needs that variable. If you make it (or its address) an input of the asm it should work as far as I can see? (Need exact code to analyse it exactly). Segher