Message ID | 1373975836-11928-2-git-send-email-haokexin@gmail.com (mailing list archive) |
---|---|
State | Accepted, archived |
Commit | e05c0e81b0628808a7490c35d1803644a18b0405 |
Delegated to: | Benjamin Herrenschmidt |
Headers | show |
On Jul 16, 2013, at 6:57 AM, Kevin Hao wrote: > For some SoC (such as the FSL BookE) even though there does have > a hardware FPU, but not all floating point instructions are > implemented. Unfortunately some versions of gcc do use these > unimplemented instructions. Then we have to enable the math emulation > to workaround this issue. It seems a little redundant to have the > support to emulate all the floating point instructions in this case. > So split the math emulation into two parts. One is for the SoC which > doesn't have FPU at all and the other for the SoC which does have the > hardware FPU and only need some special floating point instructions to > be emulated. > > Signed-off-by: Kevin Hao <haokexin@gmail.com> > --- > arch/powerpc/Kconfig | 20 ++++++++++++++++++++ > arch/powerpc/math-emu/Makefile | 24 ++++++++++++------------ > arch/powerpc/math-emu/math.c | 20 ++++++++++++++------ > 3 files changed, 46 insertions(+), 18 deletions(-) why make the split, what harm is there in just turning on the full emulation code to handle the unimplemented cases? who says what some other implementation doesn't need something that you have in CONFIG_MATH_EMULATION_FULL? Is the kernel code size really an issue? - k
On 07/22/2013 09:36:05 AM, Kumar Gala wrote: > > On Jul 16, 2013, at 6:57 AM, Kevin Hao wrote: > > > For some SoC (such as the FSL BookE) even though there does have > > a hardware FPU, but not all floating point instructions are > > implemented. Unfortunately some versions of gcc do use these > > unimplemented instructions. Then we have to enable the math > emulation > > to workaround this issue. It seems a little redundant to have the > > support to emulate all the floating point instructions in this case. > > So split the math emulation into two parts. One is for the SoC which > > doesn't have FPU at all and the other for the SoC which does have > the > > hardware FPU and only need some special floating point instructions > to > > be emulated. > > > > Signed-off-by: Kevin Hao <haokexin@gmail.com> > > --- > > arch/powerpc/Kconfig | 20 ++++++++++++++++++++ > > arch/powerpc/math-emu/Makefile | 24 ++++++++++++------------ > > arch/powerpc/math-emu/math.c | 20 ++++++++++++++------ > > 3 files changed, 46 insertions(+), 18 deletions(-) > > why make the split, what harm is there in just turning on the full > emulation code to handle the unimplemented cases? My main motivation in requesting it was to contain the increase in build time -- math-emu always stuck out to me as something that took a noticeable amount of time to build. It also reduces the increase in kernel image size. > who says what some other implementation doesn't need something that > you have in CONFIG_MATH_EMULATION_FULL? The point is to include any instructions that are known to be missing in any chip's FPU (excluding chips that don't have an FPU at all). If it is discovered that some chip is missing an instruction that we didn't account for, then we'd move that instruction from one list to the other. > Is the kernel code size really an issue? It can be when you're storing it on flash -- especially when the growth is out of control because of the need to justify pruning low-hanging fruit such as this. -Scott
diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig index 3bf72cd..7205989 100644 --- a/arch/powerpc/Kconfig +++ b/arch/powerpc/Kconfig @@ -312,6 +312,26 @@ config MATH_EMULATION such as fsqrt on cores that do have an FPU but do not implement them (such as Freescale BookE). +choice + prompt "Math emulation options" + default MATH_EMULATION_FULL + depends on MATH_EMULATION + +config MATH_EMULATION_FULL + bool "Emulate all the floating point instructions" + ---help--- + Select this option will enable the kernel to support to emulate + all the floating point instructions. If your SoC doesn't have + a FPU, you should select this. + +config MATH_EMULATION_HW_UNIMPLEMENTED + bool "Just emulate the FPU unimplemented instructions" + ---help--- + Select this if you know there does have a hardware FPU on your + SoC, but some floating point instructions are not implemented by that. + +endchoice + config PPC_TRANSACTIONAL_MEM bool "Transactional Memory support for POWERPC" depends on PPC_BOOK3S_64 diff --git a/arch/powerpc/math-emu/Makefile b/arch/powerpc/math-emu/Makefile index 8d035d2..1b46ab4 100644 --- a/arch/powerpc/math-emu/Makefile +++ b/arch/powerpc/math-emu/Makefile @@ -1,15 +1,15 @@ - -obj-$(CONFIG_MATH_EMULATION) += fabs.o fadd.o fadds.o fcmpo.o fcmpu.o \ - fctiw.o fctiwz.o fdiv.o fdivs.o \ - fmadd.o fmadds.o fmsub.o fmsubs.o \ - fmul.o fmuls.o fnabs.o fneg.o \ - fnmadd.o fnmadds.o fnmsub.o fnmsubs.o \ - fres.o fre.o frsp.o fsel.o lfs.o \ - frsqrte.o frsqrtes.o \ - fsqrt.o fsqrts.o fsub.o fsubs.o \ - mcrfs.o mffs.o mtfsb0.o mtfsb1.o \ - mtfsf.o mtfsfi.o stfiwx.o stfs.o \ - math.o fmr.o lfd.o stfd.o +math-emu-common-objs = math.o fre.o fsqrt.o fsqrts.o frsqrtes.o mtfsf.o mtfsfi.o +obj-$(CONFIG_MATH_EMULATION_HW_UNIMPLEMENTED) += $(math-emu-common-objs) +obj-$(CONFIG_MATH_EMULATION_FULL) += $(math-emu-common-objs) fabs.o fadd.o \ + fadds.o fcmpo.o fcmpu.o fctiw.o \ + fctiwz.o fdiv.o fdivs.o fmadd.o \ + fmadds.o fmsub.o fmsubs.o fmul.o \ + fmuls.o fnabs.o fneg.o fnmadd.o \ + fnmadds.o fnmsub.o fnmsubs.o fres.o \ + frsp.o fsel.o lfs.o frsqrte.o fsub.o \ + fsubs.o mcrfs.o mffs.o mtfsb0.o \ + mtfsb1.o stfiwx.o stfs.o math.o \ + fmr.o lfd.o stfd.o obj-$(CONFIG_SPE) += math_efp.o diff --git a/arch/powerpc/math-emu/math.c b/arch/powerpc/math-emu/math.c index d1ebac7..bc90162 100644 --- a/arch/powerpc/math-emu/math.c +++ b/arch/powerpc/math-emu/math.c @@ -14,6 +14,20 @@ #define FLOATFUNC(x) extern int x(void *, void *, void *, void *) +/* The instructions list which may be not implemented by a hardware FPU */ +FLOATFUNC(fre); +FLOATFUNC(frsqrtes); +FLOATFUNC(fsqrt); +FLOATFUNC(fsqrts); +FLOATFUNC(mtfsf); +FLOATFUNC(mtfsfi); + +#ifdef CONFIG_MATH_EMULATION_HW_UNIMPLEMENTED +#undef FLOATFUNC(x) +#define FLOATFUNC(x) static inline int x(void *op1, void *op2, void *op3, \ + void *op4) { } +#endif + FLOATFUNC(fadd); FLOATFUNC(fadds); FLOATFUNC(fdiv); @@ -43,8 +57,6 @@ FLOATFUNC(mcrfs); FLOATFUNC(mffs); FLOATFUNC(mtfsb0); FLOATFUNC(mtfsb1); -FLOATFUNC(mtfsf); -FLOATFUNC(mtfsfi); FLOATFUNC(lfd); FLOATFUNC(lfs); @@ -59,13 +71,9 @@ FLOATFUNC(fnabs); FLOATFUNC(fneg); /* Optional */ -FLOATFUNC(fre); FLOATFUNC(fres); FLOATFUNC(frsqrte); -FLOATFUNC(frsqrtes); FLOATFUNC(fsel); -FLOATFUNC(fsqrt); -FLOATFUNC(fsqrts); #define OP31 0x1f /* 31 */
For some SoC (such as the FSL BookE) even though there does have a hardware FPU, but not all floating point instructions are implemented. Unfortunately some versions of gcc do use these unimplemented instructions. Then we have to enable the math emulation to workaround this issue. It seems a little redundant to have the support to emulate all the floating point instructions in this case. So split the math emulation into two parts. One is for the SoC which doesn't have FPU at all and the other for the SoC which does have the hardware FPU and only need some special floating point instructions to be emulated. Signed-off-by: Kevin Hao <haokexin@gmail.com> --- arch/powerpc/Kconfig | 20 ++++++++++++++++++++ arch/powerpc/math-emu/Makefile | 24 ++++++++++++------------ arch/powerpc/math-emu/math.c | 20 ++++++++++++++------ 3 files changed, 46 insertions(+), 18 deletions(-)