On the AMD64 mailing list there was a discussion on what to set CPUTYPE to for a users system. During the discussion, the new 'native' mtune option for gcc 4.2 was brought up. This option allows gcc to automatically tune the compile for the processor that the machine is being built on for the x86 and AMD64 architectures. We need to warn users to not set CPUTYPE to 'native' in /etc/make.conf, as CPUTYPE is used to determine the value of MACHINE_CPU. But instead to use: gcc -v -x c -E -mtune=native /dev/null -o /dev/null 2>&1 | grep mtune | sed -e 's/.*mtune=//' to determine the value of CPUTYPE for their system. Fix: apply attached patch to share/examples/etc/make.conf Patch attached with submission follows:
The patch has one error, remove the second '-e' from the sed command. [linimon note: I have gone ahead and done this in the above code] -- DISCLAIMER: No electrons were mamed while sending this message. Only slightly bruised.
I found another solution to this problem by patching share/mk/bsd.cpu.mk. This patch adds code to share/mk/bsd.cpu.mk which uses gcc to change the 'native' cpu_type to the a CPUTYPE that can be used to set the correct MACHINE_CPU. Scot -- DISCLAIMER: No electrons were mamed while sending this message. Only slightly bruised.
Hello, The attached patch was unreadable, can you perhaps copy paste it in, or attach it as .txt file? Thanks! -- Kind regards, Remko Lodder ** remko@elvandar.org FreeBSD ** remko@FreeBSD.org /* Quis custodiet ipsos custodes */
On 8/28/07, Remko Lodder <remko@freebsd.org> wrote: > > Hello, > > The attached patch was unreadable, can you perhaps copy paste it in, or > attach it as .txt file? > I have to remember to add .txt when sending patches thru gmail. Scot
On 8/28/07, Scot Hetzel <swhetzel@gmail.com> wrote: > On 8/28/07, Remko Lodder <remko@freebsd.org> wrote: > > > > Hello, > > > > The attached patch was unreadable, can you perhaps copy paste it in, or > > attach it as .txt file? > > > I have to remember to add .txt when sending patches thru gmail. > I give up, no ideal as to why gnats/gmail is converting the attachment to base64 encoding, the patch is available in this post: http://lists.freebsd.org/pipermail/freebsd-current/2007-August/076503.html Scot -- DISCLAIMER: No electrons were mamed while sending this message. Only slightly bruised.
I disagree, native and ${YOUR_CPU} aren't equivalent. Here is what I get on gcc45 + core2@amd64 system: $ cc -E -v -march=core2 - </dev/null |& fgrep cc1 .../cc1 -E -quiet -v - -march=core2 $ cc -E -v -mtune=core2 - </dev/null |& fgrep cc1 .../cc1 -E -quiet -v - -mtune=core2 -march=x86-64 $ cc -E -v -march=native - </dev/null |& fgrep cc1 .../cc1 -E -quiet -v - -march=core2 -mcx16 -msahf -msse4.1 --param l1-cache-size=32 --param l1-cache-line-size=64 --param l2-cache-size=6144 -mtune=core2 $ cc -E -v -mtune=native - </dev/null |& fgrep cc1 .../cc1 -E -quiet -v - --param l1-cache-size=32 --param l1-cache-line-size=64 --param l2-cache-size=6144 -mtune=core2 -march=x86-64 And bsd.cpu.mk is oblivious about gcc43+. It will degrade native to nocona even though core2 is supported. Good example that overriding CPUTYPE is bad idea. Better leave CPUTYPE as is and populate MACHINE_CPU directly, smth like $ cat a.mk CPUTYPE = ${MACHINE_ARCH} MACHINE_CPU != echo ${MACHINE_ARCH}; ${CC} -E -dM -v -march=${CPUTYPE:S/amd64/x86-64/:S/i386/i486/} - </dev/null 2>&1 \ | awk '/SSE|MMX/ && !/MATH/ { FS="__"; gsub("_",".",$$2); print tolower($$2) }' $ make -f a.mk -V MACHINE_CPU amd64 mmx sse2 sse $ make -f a.mk -V MACHINE_CPU CPUTYPE=native amd64 sse4.1 mmx sse2 ssse3 sse sse3 $ make -f a.mk -V MACHINE_CPU CPUTYPE=native CC=/usr/bin/cc amd64 mmx sse sse2 sse3
It is very inefficient to invoke gcc, grep, and/or awk every time this makefile is processed. (In the past, we've gone to a great deal of trouble to avoid this kind of thing: http://lists.freebsd.org/pipermail/freebsd-ports/2008-July/049777.html http://lists.freebsd.org/pipermail/cvs-ports/2008-July/153224.html ) Instead, I think that comments in /etc/make.conf instructing users to how to correctly specify their CPUTYPE would be appropriate. The list of overrides for the base system compiler should be expanded, and perhaps relaxed if CC != cc. But _all_ overrides cannot be removed, and MACHINE_CPU cannot be expanded to include all features recognized by compilers from ports, because the kernel lacks the necessary support for things like SSE5/XOP+FMA4+CVT16, LWP, AVX, etc. b.
For bugs matching the following criteria: Status: In Progress Changed: (is less than) 2014-06-01 Reset to default assignee and clear in-progress tags. Mail being skipped
Keyword: patch or patch-ready – in lieu of summary line prefix: [patch] * bulk change for the keyword * summary lines may be edited manually (not in bulk). Keyword descriptions and search interface: <https://bugs.freebsd.org/bugzilla/describekeywords.cgi>
Created attachment 257964 [details] make.conf.diff ^Triage: rebase patch.