powerpc/spe: tweak glibc/eglibc/uclibc combinations
PATCH 1/2 powerpc/spe: tweak glibc/eglibc/uclibc combinations
Since (e)glibc 2.19+ PowerPC SPE changed ABI conventions for sanity. In the good old days hard-float was used, it sounds good on paper until you account for the e500v1/v2 differences, with v1 being a fpu that does only single-precision math.
So the ABI changed to make it soft-float and the libc (or libm) handle that deficiency in software if required, making old precompiled binaries incompatible (of no concern for buildroot since we've got no precompiled binary blobs for SPE in packages).
For uClibc the support is for hard-float so disable it for soft since it doesn't do the fallback trickery.
So do what's appropiate for each combination:
H-F S-F
eglibc 2.18 x
eglibc 2.19 x glibc 2.19 x glibc 2.20 x
uclibc x
Eventually hard-float should just go away, i'm keeping it for compatibility purposes.
--- package/glibc/Config.in | 4 ++-- toolchain/toolchain-buildroot/Config.in | 3 ++- 2 files changed, 4 insertions(+), 3 deletions(-)
package/glibc/Config.in +2 -2
@@ -11,11 +11,11 @@ choice config BR2_EGLIBC_VERSION_2_18 bool "2.18-svnr23787" + depends on !BR2_POWERPC_SOFT_FLOAT config BR2_EGLIBC_VERSION_2_19 bool "2.19-svnr25243" - # Build breakage - depends on !BR2_powerpc_SPE + depends on !BR2_powerpc_SPE || (BR2_powerpc_SPE && BR2_POWERPC_SOFT_FLOAT) endchoice
toolchain/toolchain-buildroot/Config.in +2 -1
@@ -33,6 +33,7 @@ config BR2_TOOLCHAIN_BUILDROOT_UCLIBC BR2_mips || BR2_mipsel || BR2_mips64 || BR2_mips64el || \ BR2_powerpc || BR2_sh || BR2_sparc || BR2_xtensa || \ BR2_x86_64 + depends on !BR2_powerpc_SPE || (BR2_powerpc_SPE && !BR2_POWERPC_SOFT_FLOAT) help This option selects uClibc as the C library for the cross-compilation toolchain.
@@ -67,7 +68,7 @@ config BR2_TOOLCHAIN_BUILDROOT_GLIBC BR2_microblaze depends on BR2_USE_MMU depends on !BR2_PREFER_STATIC_LIB - depends on !BR2_powerpc_SPE + depends on !BR2_powerpc_SPE || (BR2_powerpc_SPE && BR2_POWERPC_SOFT_FLOAT) select BR2_TOOLCHAIN_USES_GLIBC # our glibc.mk enables RPC support select BR2_TOOLCHAIN_HAS_NATIVE_RPC
Signature · 1 line
1.8.5.5
PATCH 2/2 arch/powerpc: default to soft-float for SPE ABI
It really is the way forward so switch the default.
--- arch/Config.in.powerpc | 1 + 1 file changed, 1 insertion(+)
arch/Config.in.powerpc +1 -0
@@ -141,6 +141,7 @@ endchoice config BR2_POWERPC_SOFT_FLOAT bool "Use soft-float" + default y if BR2_powerpc_SPE select BR2_SOFT_FLOAT help If your target CPU does not have a Floating Point Unit (FPU)
Signature · 1 line
1.8.5.5
PATCH 1/2 powerpc/spe: tweak glibc/eglibc/uclibc combinations
In reply to Gustavo ZacariasDear Gustavo Zacarias,
Quoted 9 lines · Gustavo Zacarias
Since (e)glibc 2.19+ PowerPC SPE changed ABI conventions for sanity. In the good old days hard-float was used, it sounds good on paper until you account for the e500v1/v2 differences, with v1 being a fpu that does only single-precision math.
So the ABI changed to make it soft-float and the libc (or libm) handle that deficiency in software if required, making old precompiled binaries incompatible (of no concern for buildroot since we've got no precompiled binary blobs for SPE in packages).
I'm not sure to fully understand: is PowerPC really using *soft-float* as in all floating point operations are emulated in software? Or is it a difference similar to the difference between hardfp and softfp on ARM (both use the FPU and floating point instructions, it's only a different ABI, where hardfp passes floating point arguments in floating point registers, while softfp passes floating point arguments in integer registers).
Thanks,
Thomas
Signature · 3 lines
Thomas Petazzoni, CTO, Free Electrons Embedded Linux, Kernel and Android engineering http://free-electrons.com
PATCH 1/2 powerpc/spe: tweak glibc/eglibc/uclibc combinations
Quoted 7 lines · Thomas Petazzoni
I'm not sure to fully understand: is PowerPC really using *soft-float* as in all floating point operations are emulated in software? Or is it a difference similar to the difference between hardfp and softfp on ARM (both use the FPU and floating point instructions, it's only a different ABI, where hardfp passes floating point arguments in floating point registers, while softfp passes floating point arguments in integer registers).
It's similar to the ARM scenario, yes.
In the SPE ABI there are no FPU registers at all, it's just an extension of GP-ones (say 32+32 bit for DP/E500v2).
In libc-land uclibc uses the 'old' ABI (direct hard float), and so did eglibc <= 2.18 (glibc i think never got that part officially). To avoid re-explaining: http://www.eglibc.org/archives/patches/msg01291.html
I hope it's clear enough :)
Regards.