On Thu, May 14, 2020 at 02:27:20AM +0900, Masahiro Yamada wrote: > On Thu, May 14, 2020 at 1:13 AM Tom Rini <tr...@konsulko.com> wrote: > > > > On Thu, May 14, 2020 at 01:05:37AM +0900, Masahiro Yamada wrote: > > > Hi Tom, > > > > > > On Wed, May 13, 2020 at 11:42 PM Tom Rini <tr...@konsulko.com> wrote: > > > > > > > > On Tue, May 12, 2020 at 11:04:38PM -0400, Tom Rini wrote: > > > > > On Mon, May 11, 2020 at 09:08:03PM +0200, Heinrich Schuchardt wrote: > > > > > > On 5/11/20 8:40 PM, Tom Rini wrote: > > > > > > > On Sun, May 10, 2020 at 10:12:07PM +0900, Masahiro Yamada wrote: > > > > > > >> On Sun, May 10, 2020 at 12:12 AM Heinrich Schuchardt > > > > > > >> <xypron.g...@gmx.de> wrote: > > > > > > >>> > > > > > > >>> GCC recognizes /* fallthrough */ if -Wimplicit-fallthrough=3 is > > > > > > >>> enabled. > > > > > > >> > > > > > > >> FYI. > > > > > > >> > > > > > > >> Linux decided to not use /* fallthrough */ any more > > > > > > >> because Clang does not recognize it. > > > > > > >> > > > > > > >> __attribute__((__fallthrough__)) is supported > > > > > > >> by both Clang and recent GCC. > > > > > > In fact Linux has a define: > > > > > > > > > > > > include/linux/compiler_attributes.h:200:# define fallthrough > > > > > > __attribute__((__fallthrough__)) > > > > > > > > > > > > And in the code you would use > > > > > > > > > > > > case foo: > > > > > > fallthrough; > > > > > > case bar: > > > > > > > > > > > > But the Linux kernel still has a lot of lines with > > > > > > > > > > > > /* fallthrough */ > > > > > > > > > > > > Documentation/process/deprecated.rst: > > > > > > > > > > > > <cite> > > > > > > As there have been a long list of flaws `due to missing "break" > > > > > > statements <https://cwe.mitre.org/data/definitions/484.html>`_, we > > > > > > no > > > > > > longer allow implicit fall-through. In order to identify intentional > > > > > > fall-through cases, we have adopted a pseudo-keyword macro > > > > > > "fallthrough" > > > > > > which expands to gcc's extension `__attribute__((__fallthrough__)) > > > > > > <https://gcc.gnu.org/onlinedocs/gcc/Statement-Attributes.html>`_. > > > > > > (When > > > > > > the C17/C18 `[[fallthrough]]` syntax is more commonly supported by > > > > > > C > > > > > > compilers, static analyzers, and IDEs, we can switch to using that > > > > > > syntax for the macro pseudo-keyword.) > > > > > > </cite> > > > > > > > > > > > > Using the attribute is not standard C and not any better than using > > > > > > the > > > > > > comment. The real target is the C17 syntax. > > > > > > > > > > > > >> > > > > > > >> > > > > > > >> Linux is now doing treewide conversion > > > > > > >> from /* fallthrough */ to 'fallthrough;'. > > > > > > >> > > > > > > >> See include/linux/compiler_attributes.h in Linux. > > > > > > >> > > > > > > >> I do not know if U-Boot wants to align with it. > > > > > > >> (up to Tom ?) > > > > > > > > > > > > > > A re-sync on the compiler headers again and making use of this > > > > > > > sounds > > > > > > > like a good idea, yes. > > > > > > > > > > > > > > > > > > > We should enable -Wimplicit-fallthrough like the kernel does. This > > > > > > defaults to -Wimplicit-fallthrough=3 and is happy with both the > > > > > > comment > > > > > > as well as with the attribute. > > > > > > > > > > > > @Tom: > > > > > > Will you update the compiler headers within this release cycle? > > > > > > Otherwise we should take the patch as is to get us closer to the > > > > > > -Wimplicit-fallthrough target. > > > > > > > > > > I'm not going to update it for this release cycle. I've done the > > > > > initial import and build and there's some fairly large changes related > > > > > to inlining that I want to look at harder to see if we can/should do > > > > > something about (I don't want to derail this thread, I'll start > > > > > another). But it's very far from zero size change and given the > > > > > inline > > > > > changes I think it'll need real testing. > > > > > > > > > > And since the kernel isn't making a huge use yet of fallthrough; we > > > > > can > > > > > afford to look a little harder at things. > > > > > > > > I think I've figured out the inline issue which is that we need > > > > scripts/Kconfig.include from the kernel, CC_HAS_ASM_INLINE Kconfig > > > > option, and re-sync with Kconfiglib, but that's still going to be enough > > > > stuff that I don't think it's good to pull in at -rc2. > > > > > > > > > > > > > I do not get how 'asm inline' support is related > > > to this topic. > > > > > > GCC 9 started to support 'asm inline' for the better inlining heuristic. > > > The kernel uses a bunch of inline assembly > > > that is not as expensive as it looks. > > > > > > As GCC is agnostic about the real cost of inline assembly, > > > 'asm inline' is a good hint if people know the real cost is quite small. > > > Then, GCC will be able to inline more functions. > > > > > > I do not know how important it is for U-Boot, though. > > > > > > What is causing you a trouble? > > > > So, it turns out that while we do want to grab the changes so that we > > can have CC_HAS_ASM_INLINE via Kconfig, it's not "it". What I see for > > virtually every board (with gcc-9.3 from kernel.org) is something like: > > rock960-rk3399 : all -8 rodata -4 spl/u-boot-spl:all +992 > > spl/u-boot-spl:text +992 text -4 > > u-boot: add: 67/-9, grow: 74/-92 bytes: 5072/-4928 (144) > > function old new > > delta > > static._compare_and_overwrite_entry - 348 > > +348 > > menu_interactive_choice - 288 > > +288 > > hex2bin - 200 > > +200 > > __fswab64 - 176 > > +176 > > __fswab32 - 144 > > +144 > > sdhci_reset - 136 > > +136 > > dwmci_fifo_ready - 124 > > +124 > > fdt_offset_ptr_ - 120 > > +120 > > menu_items_iter - 108 > > +108 > > generic_fls - 100 > > +100 > > fdt_set_totalsize - 96 > > +96 > > static.generic_fls - 84 > > +84 > > > OK, these functions previously disappeared because all > of the function call-sites were inlined. > > If you resync <linux/compier*.h> with latest Linux, > they are not necessarily inlined. > > > > > In current U-Boot, 'static inline' is actually replaced with > __attribute__((always_inline)). > So, inlining is forcible. > > See the code. > > > include/linux/compiler-gcc.h > > #if !defined(CONFIG_ARCH_SUPPORTS_OPTIMIZED_INLINING) || \ > !defined(CONFIG_OPTIMIZE_INLINING) || (__GNUC__ < 4) > #define inline inline __attribute__((always_inline)) notrace > #define __inline__ __inline__ __attribute__((always_inline)) notrace > #define __inline __inline __attribute__((always_inline)) notrace > > > > > In Linux, the following commits stopped doing that. > (both my commits) > > ac7c3e4ff401b304489a031938dbeaab585bfe0a > 889b3c1245de48ed0cacf7aebb25c489d3e4a3e9 > > > Now, 'inline' is just a compiler hint. > The compiler does the best judge > whether to inline the function or not.
Ah, OK. And the kernel is uninterested in bringing back forcing inlineing for size reasons as it's gone for LTO instead of ffunction-sections/fdata-sections and discarding sections (and we're in turn a ways off from that as we need to move to gcc linking) I assume. Thanks! -- Tom
signature.asc
Description: PGP signature