For those playing on the bleeding edge with optimise_gcc.mk,
disable -ffast-math for perl. NetBSD/i386 perl5.8.7 reports
'49/49!=1' as true under -ffast-math
- Add Infineon ADM5120
- Fix optimize_gcc.mk for qt3-libs, and reorder so CPUFLAGS_PKGBASE is
left defined and testable for 'make show-vars' and the like
optimize_gcc.mk needs to know the name of the package that is
currently being built. Unfortunately with pkgsrc there is no way
to do this, and its previous attempts to derive PKGNAME (if not
set) from DISTNAME broke the php usage ext.mk file. Use a temp
variable to do the same thing without breaking other pkgsrc
assumptions
- Added linux "Mobile Intel(R) Pentium(R) 4 - M ", from pancake
- Add a dist target to simplify updating freshmeat copy
- Change feedback email to abs@absd.org
Add Linux 'Pentium III (Katmai)' - from Toni Viemero
Remove some outdated gcc references from manpage
Clarify the mk examples are BSD makefile specific
Add a basic Makefile for out-of-pkgsrc usage
firefox-gtk2 has severe issues (radio buttons on the download preferences
are missing, and extensions break) if compiled with -ffast-math. Oh my
aching head.
- Shuffle the various cpuflags.<OS> to group more common lines
- Use nawk on solaris [Jonathan A. Kollasch PR pkg/26572]
- Add Linux 'Intel(R) Pentium(R) 4 CPU'
- Add NetBSD/i386 'AMD Athlon Model 4 (Thunderbird)' [Rob Quinn]
- Add NetBSD/macppc 604e [John Klos]
- Add NetBSD/chocolate [Al]
- all:
Implement 'cpuflags -v' which indicates the values used to determine flags
Sync the latest fixup options for older gccs
- cpuflags.NetBSD:
Use -march=pentium3m and -march=pentium4m for gcc 3.4.x
Set -mcpu=hypersparc for RT620/625 and -mcpu=cypress for L64811
- cpuflags.Linux:
Handle two different Celeron cases (thanks to Marc Recht & Jeremy C. Reed)
NetBSD/Linux:
athlon-4, athlon-mp, athlon-tbird, and athlon-xp should fall back to
athlon in gcc3, and only pentiumpro in 2.95
Linux:
Identify "AMD Athlon(tm) XP "* - thanks to Sebastian Prause
Use -dumpversion with gcc rather than mangling -v output. Suggested
by mrg. Located copy of egcs to confirm output still correctly
mangled in that case :)
- Do nothing on x86_64 (for now). Thanks to Wolfgang S. Rupprecht
- Digital AlphaPC 164SX. Thanks to HATANO Hiromichi.
- Intel Pentium/MMX & Unrecognised Athlon XP. Thanks to Kimmo Suominen
- Mips update for gcc 3.1 and later. Thanks to Andreas Yankopolus
- Handle converting flags for old gcc versions recursively. Most
excessive case would be using cpuflags on a k6-3 win gcc 2.89 which
would map -march=k6-3 -> -march=k6 -> -march=pentium -> -march=i486
"Athlon MP" nor a "Pentium Pro". Correct this by testing for the special
cases first rather than trying to be clever in the generic test.
Bump package version to 0.58
Add optimisations for specific Athlon models (recht@)
Avoid setting flags if DBG is set to -Os (to reduce size impact
when building contents of distrib on at least 1.6.x i386)
- If !defined BSD_PKG_MK and DBG is set to -Os, do not alter COPTS.
This stops setting of -march=XXX causing overflowing of install media
- disable -fomit-frame-pointer for lua and koffice
Handle hw.model frmatting from hp300
On gcc3 -frename-registers still causes problems for xdm with _some_
-march=XXX values. Next time I might just listen to fredb first...
Add mozilla and pango to the list of packages for which
-frename-registers triggers bug in the binutils shipped with 1.6.x
(Thanks to Frederick Bruckman for identifying the issue)
Treat AMD Duron as Athlon
Replace PKG_EXCLUDE_O3 with PKG_EXCLUDE_RENAME_REGISTERS as its
specifically the -frename-registers that causes problems. Set for ORBit
- Detect fpuless mac68k boxes and add -msoft-float
- Check ${USETOOLS} to avoid setting gcc3 flags when compiling NetBSD
userland with built in tools, but with gcc3 installed
- Default to -O3 again for all except perl5 and perl58 under gcc3
Convert -march=athlon for older gccs to -march=pentiumpro.
Addresses PR pkg/21766 by Marc Recht.
XXX: the conversion mechanism is suboptimal, it can't handle recursive
conversions.
- Add some additional x86 flags for gcc 3.x
- Allow the path to cc to be specified as an argument to cpuflags
- Use both the previous to handle the gcc3 package
- Add an 'unsupported' optimize_gcc.mk which... to quote its comments:
"This file is 'experimental' - which is doublespeak for unspeakably
ugly, and probably quite broken by design.
The intention is to pass additional flags to gcc to further optimise
generated code. It _will_ make it impossible to debug, may fail to
compile some code, and even generate curdled binaries. It is completely
unsupported. Any questions should be directed to <abs@netbsd.org>.
cpuflags.mk now sets CPU_FLAGS and CPU_DIR in .MAKEFLAGS to
avoid having to call cpuflags in every sub-make. Reduces a
'make clean' in /usr/src on a celeron 850 from 90 seconds to
30 seconds. What is impressive is that I managed to leave it
that bad without being bricked...
other friendly people on port-macppc.
While I'm here rework to check hw.machine_arch first and then hw.model, so
we cna handle overlapping model matches between hw.machine_arch. We've not
hit any yet, but this is the right thing to do :)
Determine the appropriate flags for the Alpha 433a I now have.
(Insert appropriate grumble about there being no sysctl way to tell
the cputype on an Alpha)
Update SA-110 in Risc PC with problem memory bus switch:
- Pickup post 1.5 cats, dnard, hpcarm, or netwinder as OK.
- If in 1.5 check for footbridge0 or ofw0 as OK.
Thanks to Chris Gilbert for details.
