On Thu, Oct 15, 2015 at 3:02 AM, Phil Blundell <[email protected]> wrote: > On Wed, 2015-10-14 at 23:37 -0700, Khem Raj wrote: >> > On Oct 14, 2015, at 2:38 PM, Andre McCurdy <[email protected]> >> > wrote: >> > >> > I'm seeing what looks like a gcc bug when building syslinux with >> > certain combinations of TUNE_CCARGS. A specific combination which >> > fails is: >> > >> > TUNE_CCARGS = " -m32 -march=core2 -mtune=core2 -msse3 -mfpmath=sse >> > -fschedule-insns” >> >> >> and why would you use -fschedule-insns on something like ia32 with so >> few general purpose regs >> at disposal ? > > -fschedule-insns is enabled by default at -O2 anyway so I'm slightly > surprised that this is making a difference. As far as I know it is > entirely reasonable to use that option on x86: there is still scope for > instruction reordering even though the number of GP regs is small. >
There are other options that are also involved here, anyway what I was pointing to was that insn pass happens before machine regs are assigned so it may not always work and cause this kind of ICEs especially on machines with smaller number of GPs >> applying any insn schedule opts before machine registers are assigned >> may not always work. > > I don't really understand this comment. Are you saying that you think > only -fschedule-insns2 should be used, not -fschedule-insns? insn2 is safer since its run after register allocation. so you wont have the register pressure issue. > > p. > -- _______________________________________________ Openembedded-core mailing list [email protected] http://lists.openembedded.org/mailman/listinfo/openembedded-core
