| Message ID | 20171215232530.9388-1-stefan.froberg@petroprogram.com |
|---|---|
| State | Rejected |
| Delegated to: | Thomas Petazzoni |
| Headers | show |
| Series | [1/1] gcc: Support configure option --with-linker-hash-style | expand |
Hello, On Sat, 16 Dec 2017 01:25:30 +0200, Stefan Fröberg wrote: > This will set the default hash style that GCC will > always use during linking. > > Signed-off-by: Stefan Fröberg <stefan.froberg@petroprogram.com> Could you explain the motivation for configuring the default hash style? If "gnu" gives the best application startup time, why would we use a different one? Generally, the commit log lacks a bit of background, i.e the "why" we would want such a change. Thanks! Thomas
Hi Thomas Well, I was just worried that if setting the linking hash style to stone (aka "gnu") then could it maybe break builds in some other archs than x86, x86_64 and ARM ? So I played carefull and made it configurable by choice. Im not sure but I think that option has been hiding in GCC since 4.x and buildroot lowest is 4.9. So if it's okay I can cook another patch where it just set's it in stone without choice ? -S- Thomas Petazzoni kirjoitti 16.12.2017 klo 17:23: > Hello, > > On Sat, 16 Dec 2017 01:25:30 +0200, Stefan Fröberg wrote: >> This will set the default hash style that GCC will >> always use during linking. >> >> Signed-off-by: Stefan Fröberg <stefan.froberg@petroprogram.com> > Could you explain the motivation for configuring the default hash > style? If "gnu" gives the best application startup time, why would we > use a different one? > > Generally, the commit log lacks a bit of background, i.e the "why" we > would want such a change. > > Thanks! > > Thomas
Hi Thomas, I'm looking at this oldest patch in the patchwork... So, at the moment we use both hash (sysv and gnu) by default for toolchains with binutils >= 2.30, except for mips architecure which still use sysv hash. See [1]: # Enable gnu hash only on GNU targets, but not mips mips*-*-*) enable_default_hash_style=sysv ;; *-*-gnu* | *-*-linux* | *-*-nacl*) enable_default_hash_style=both ;; *) enable_default_hash_style=sysv ;; gnu hash style was introduced by [2] (binutils 2.18) The gnu hash style is enabled by default since binutils 2.30 [3] along with sysv hash style. If --hash-style option is supported by the compiler, glibc use both hash (sysv and gnu) [4]. But this doesn't mean that all binaries will be build with gnu style hash if binutils < 2.30. So, I'm not sure we want to enable only gnu style for all binaries on the target (especially the libc). Binaries that contains only gnu hash style are incompatible with binaries that contains only sysv hash style. The default for the choice should be BR2_GCC_HASH_STYLE_BOTH instead of BR2_GCC_HASH_STYLE_GNU if we want to make this option configurable. Best regards, Romain [1] https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=blob;f=ld/configure;h=e46248cd089e71323ed20bc33ab0efdde526a2bc;hb=refs/heads/binutils-2_31-branch#l15884 [2] https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=fdc90cb46b0f96dd4444ec3d126c87de75fb6e6b [3] https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=2760f24c4942853eac7b921e4b8843d57a602654 [4] https://sourceware.org/git/?p=glibc.git;a=commit;h=871b91589bf4f6dfe19d5987b0a05bd7cf936ecc Le 17/12/2017 à 15:30, Stefan Fröberg a écrit : > Hi Thomas > > Well, I was just worried that if setting the linking hash style to stone > (aka "gnu") then could > it maybe break builds in some other archs than x86, x86_64 and ARM ? > > So I played carefull and made it configurable by choice. > > Im not sure but I think that option has been hiding in GCC since 4.x and > buildroot lowest is 4.9. > So if it's okay I can cook another patch where it just set's it in stone > without choice ? > > -S- > > Thomas Petazzoni kirjoitti 16.12.2017 klo 17:23: >> Hello, >> >> On Sat, 16 Dec 2017 01:25:30 +0200, Stefan Fröberg wrote: >>> This will set the default hash style that GCC will >>> always use during linking. >>> >>> Signed-off-by: Stefan Fröberg <stefan.froberg@petroprogram.com> >> Could you explain the motivation for configuring the default hash >> style? If "gnu" gives the best application startup time, why would we >> use a different one? >> >> Generally, the commit log lacks a bit of background, i.e the "why" we >> would want such a change. >> >> Thanks! >> >> Thomas > > _______________________________________________ > buildroot mailing list > buildroot@busybox.net > http://lists.busybox.net/mailman/listinfo/buildroot >
On 09/11/2018 00:49, Romain Naour wrote: > Hi Thomas, > > I'm looking at this oldest patch in the patchwork... Great! > > So, at the moment we use both hash (sysv and gnu) by default for toolchains with > binutils >= 2.30, except for mips architecure which still use sysv hash. > > See [1]: > # Enable gnu hash only on GNU targets, but not mips > mips*-*-*) enable_default_hash_style=sysv ;; > *-*-gnu* | *-*-linux* | *-*-nacl*) enable_default_hash_style=both ;; > *) enable_default_hash_style=sysv ;; > > gnu hash style was introduced by [2] (binutils 2.18) > > The gnu hash style is enabled by default since binutils 2.30 [3] along with sysv > hash style. To be exact: since 2.30, binutils has an option for the default hash style, and that default is "both" (except for mips), while before the default was "sysv". > If --hash-style option is supported by the compiler, ... which it is for all gcc versions we support... > glibc use both hash (sysv > and gnu) [4]. But this doesn't mean that all binaries will be build with gnu > style hash if binutils < 2.30. ... or if it is mips. However, if GCC is configured with --with-linker-hash-style=foo, then all binaries *will* actually be built with that hash style, no? > So, I'm not sure we want to enable only gnu style for all binaries on the target > (especially the libc). But if I understand correctly, glibc overrides it anyway and always sets "both"? So for glibc, at least, there shouldn't be a problem? As far as I understand, there shouldn't be a problem because all libraries and executables *will* be built with the same hash style. There may be three problems though: 1. If binutils defaults to "sysv" for mips, perhaps the GNU hash style is broken there? So perhaps we should disable the option for mips? 2. Some packages may (actually do) override the default hash style. So if you select hash style sysv in gcc, and use a libc which doesn't override it, and then try to link a program that itself overrides to hash style gnu, it will fail. 3. Similar for binary packages: we don't build them, so we don't control the hash style. Probably, though, binary packages will use hash style both. > Binaries that contains only gnu hash style are > incompatible with binaries that contains only sysv hash style. > > The default for the choice should be BR2_GCC_HASH_STYLE_BOTH instead of > BR2_GCC_HASH_STYLE_GNU if we want to make this option configurable. Or we could just bite the bullet and see how it goes wrong :-) It makes some kind of sense to make it configurable for now, so that there is a way out if there are a few packages that fail with the wrong hash style. But then at least we should randomize the options in the autobuilders. The good thing is that any problems will be immediately visible at link time (I think at least that that's the case; I hope it's not at dynamic load time that the problem occurs... And of course, dlopen()'ed libraries will only have runtime problems.) Regards, Arnout > > Best regards, > Romain > > [1] > https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=blob;f=ld/configure;h=e46248cd089e71323ed20bc33ab0efdde526a2bc;hb=refs/heads/binutils-2_31-branch#l15884 > [2] > https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=fdc90cb46b0f96dd4444ec3d126c87de75fb6e6b > [3] > https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=2760f24c4942853eac7b921e4b8843d57a602654 > [4] > https://sourceware.org/git/?p=glibc.git;a=commit;h=871b91589bf4f6dfe19d5987b0a05bd7cf936ecc > > Le 17/12/2017 à 15:30, Stefan Fröberg a écrit : >> Hi Thomas >> >> Well, I was just worried that if setting the linking hash style to stone >> (aka "gnu") then could >> it maybe break builds in some other archs than x86, x86_64 and ARM ? >> >> So I played carefull and made it configurable by choice. >> >> Im not sure but I think that option has been hiding in GCC since 4.x and >> buildroot lowest is 4.9. >> So if it's okay I can cook another patch where it just set's it in stone >> without choice ? >> >> -S- >> >> Thomas Petazzoni kirjoitti 16.12.2017 klo 17:23: >>> Hello, >>> >>> On Sat, 16 Dec 2017 01:25:30 +0200, Stefan Fröberg wrote: >>>> This will set the default hash style that GCC will >>>> always use during linking. >>>> >>>> Signed-off-by: Stefan Fröberg <stefan.froberg@petroprogram.com> >>> Could you explain the motivation for configuring the default hash >>> style? If "gnu" gives the best application startup time, why would we >>> use a different one? >>> >>> Generally, the commit log lacks a bit of background, i.e the "why" we >>> would want such a change. >>> >>> Thanks! >>> >>> Thomas >> >> _______________________________________________ >> buildroot mailing list >> buildroot@busybox.net >> http://lists.busybox.net/mailman/listinfo/buildroot >> > > _______________________________________________ > buildroot mailing list > buildroot@busybox.net > http://lists.busybox.net/mailman/listinfo/buildroot >
Hi Arnout, Le 09/11/2018 à 22:35, Arnout Vandecappelle a écrit : > > > On 09/11/2018 00:49, Romain Naour wrote: >> Hi Thomas, >> >> I'm looking at this oldest patch in the patchwork... > > Great! > >> >> So, at the moment we use both hash (sysv and gnu) by default for toolchains with >> binutils >= 2.30, except for mips architecure which still use sysv hash. >> >> See [1]: >> # Enable gnu hash only on GNU targets, but not mips >> mips*-*-*) enable_default_hash_style=sysv ;; >> *-*-gnu* | *-*-linux* | *-*-nacl*) enable_default_hash_style=both ;; >> *) enable_default_hash_style=sysv ;; >> >> gnu hash style was introduced by [2] (binutils 2.18) >> >> The gnu hash style is enabled by default since binutils 2.30 [3] along with sysv >> hash style. > > To be exact: since 2.30, binutils has an option for the default hash style, and > that default is "both" (except for mips), while before the default was "sysv". Indeed. > > >> If --hash-style option is supported by the compiler, > > ... which it is for all gcc versions we support... > >> glibc use both hash (sysv >> and gnu) [4]. But this doesn't mean that all binaries will be build with gnu >> style hash if binutils < 2.30. > > ... or if it is mips. > > However, if GCC is configured with --with-linker-hash-style=foo, then all > binaries *will* actually be built with that hash style, no? I'm not sure if gcc's "--with-linker-hash-style=foo" (added in 4.7.0 [1]) win against "-Wl,--hash-style=bar" added by glibc. (This also mean --with-linker-hash-style needs gcc >= 4.7.0). "Prepend --hash-style=LINKER_HASH_STYLE to whatever link_spec we had before." We should have : "--hash-style=foo --hash-style=bar" on the linker command line. ("bar" should be used in this case I believe) [1] https://github.com/gcc-mirror/gcc/commit/3cb9bbfa927aa187048534f9069202c017a78e38 > > >> So, I'm not sure we want to enable only gnu style for all binaries on the target >> (especially the libc). > > But if I understand correctly, glibc overrides it anyway and always sets > "both"? So for glibc, at least, there shouldn't be a problem? This must be verified. > > > As far as I understand, there shouldn't be a problem because all libraries and > executables *will* be built with the same hash style. There may be three > problems though: > > 1. If binutils defaults to "sysv" for mips, perhaps the GNU hash style is broken > there? So perhaps we should disable the option for mips? Well, we don't have any explanation in the commit log adding the --enable-default-hash-style option about mips. We don't know why we need to use sysv on mips. > > 2. Some packages may (actually do) override the default hash style. So if you > select hash style sysv in gcc, and use a libc which doesn't override it, and > then try to link a program that itself overrides to hash style gnu, it will fail. Indeed. IIUC it should be a runtime problem with the dynamic loader. I haven't tested this patch so far... > > 3. Similar for binary packages: we don't build them, so we don't control the > hash style. Probably, though, binary packages will use hash style both. Maybe we have to check each binary package (ex opengl Mali). > > >> Binaries that contains only gnu hash style are >> incompatible with binaries that contains only sysv hash style. >> >> The default for the choice should be BR2_GCC_HASH_STYLE_BOTH instead of >> BR2_GCC_HASH_STYLE_GNU if we want to make this option configurable. > > Or we could just bite the bullet and see how it goes wrong :-) :) > > It makes some kind of sense to make it configurable for now, so that there is a > way out if there are a few packages that fail with the wrong hash style. But > then at least we should randomize the options in the autobuilders. The good > thing is that any problems will be immediately visible at link time (I think at > least that that's the case; I hope it's not at dynamic load time that the > problem occurs... And of course, dlopen()'ed libraries will only have runtime > problems.) Actually, it's not clear to me about the need of using hash sysv only or gnu only, why using both hash is not sufficient? Best regards, Romain > > > Regards, > Arnout > > >> >> Best regards, >> Romain >> >> [1] >> https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=blob;f=ld/configure;h=e46248cd089e71323ed20bc33ab0efdde526a2bc;hb=refs/heads/binutils-2_31-branch#l15884 >> [2] >> https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=fdc90cb46b0f96dd4444ec3d126c87de75fb6e6b >> [3] >> https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=2760f24c4942853eac7b921e4b8843d57a602654 >> [4] >> https://sourceware.org/git/?p=glibc.git;a=commit;h=871b91589bf4f6dfe19d5987b0a05bd7cf936ecc >> >> Le 17/12/2017 à 15:30, Stefan Fröberg a écrit : >>> Hi Thomas >>> >>> Well, I was just worried that if setting the linking hash style to stone >>> (aka "gnu") then could >>> it maybe break builds in some other archs than x86, x86_64 and ARM ? >>> >>> So I played carefull and made it configurable by choice. >>> >>> Im not sure but I think that option has been hiding in GCC since 4.x and >>> buildroot lowest is 4.9. >>> So if it's okay I can cook another patch where it just set's it in stone >>> without choice ? >>> >>> -S- >>> >>> Thomas Petazzoni kirjoitti 16.12.2017 klo 17:23: >>>> Hello, >>>> >>>> On Sat, 16 Dec 2017 01:25:30 +0200, Stefan Fröberg wrote: >>>>> This will set the default hash style that GCC will >>>>> always use during linking. >>>>> >>>>> Signed-off-by: Stefan Fröberg <stefan.froberg@petroprogram.com> >>>> Could you explain the motivation for configuring the default hash >>>> style? If "gnu" gives the best application startup time, why would we >>>> use a different one? >>>> >>>> Generally, the commit log lacks a bit of background, i.e the "why" we >>>> would want such a change. >>>> >>>> Thanks! >>>> >>>> Thomas >>> >>> _______________________________________________ >>> buildroot mailing list >>> buildroot@busybox.net >>> http://lists.busybox.net/mailman/listinfo/buildroot >>> >> >> _______________________________________________ >> buildroot mailing list >> buildroot@busybox.net >> http://lists.busybox.net/mailman/listinfo/buildroot >> > _______________________________________________ > buildroot mailing list > buildroot@busybox.net > http://lists.busybox.net/mailman/listinfo/buildroot >
diff --git a/package/gcc/Config.in.host b/package/gcc/Config.in.host index 70cce0a5c5..b9391392e1 100644 --- a/package/gcc/Config.in.host +++ b/package/gcc/Config.in.host @@ -152,3 +152,30 @@ config BR2_GCC_ENABLE_GRAPHITE comment "graphite support needs gcc >= 5.x" depends on !BR2_TOOLCHAIN_GCC_AT_LEAST_5 + +choice + prompt "Hash style used during linking" + default BR2_GCC_HASH_STYLE_GNU + help + This tells GCC to pass --hash-style=choice option to the + linker for all final links where choice can be one of + "gnu", "sysv" or "both". + Using "gnu" should give best application startup time. + +config BR2_GCC_HASH_STYLE_GNU + bool "gnu" + help + For new style GNU ".gnu.hash" section. + +config BR2_GCC_HASH_STYLE_SYSV + bool "sysv" + help + For classic ELF ".hash" section. + +config BR2_GCC_HASH_STYLE_BOTH + bool "both" + help + For both the classic ELF ".hash" and + new style GNU ".gnu.hash". + +endchoice diff --git a/package/gcc/gcc.mk b/package/gcc/gcc.mk index 27fc1e987c..51bb22d0e9 100644 --- a/package/gcc/gcc.mk +++ b/package/gcc/gcc.mk @@ -183,6 +183,18 @@ else HOST_GCC_COMMON_CONF_OPTS += --without-isl --without-cloog endif +ifeq ((BR2_GCC_HASH_STYLE_GNU),y) +HOST_GCC_COMMON_CONF_OPTS += --with-linker-hash-style=gnu +endif + +ifeq ((BR2_GCC_HASH_STYLE_SYSV),y) +HOST_GCC_COMMON_CONF_OPTS += --with-linker-hash-style=sysv +endif + +ifeq ((BR2_GCC_HASH_STYLE_BOTH),y) +HOST_GCC_COMMON_CONF_OPTS += --with-linker-hash-style=both +endif + ifeq ($(BR2_arc)$(BR2_or1k),y) HOST_GCC_COMMON_DEPENDENCIES += host-flex host-bison endif
This will set the default hash style that GCC will always use during linking. Signed-off-by: Stefan Fröberg <stefan.froberg@petroprogram.com> --- package/gcc/Config.in.host | 27 +++++++++++++++++++++++++++ package/gcc/gcc.mk | 12 ++++++++++++ 2 files changed, 39 insertions(+)