[Buildroot] [PATCH] package/libnss: fix build failure on arm32 arch not armv7

[Giulio Benetti]

Hi Thomas,

On 3/25/20 3:45 PM, Thomas Petazzoni wrote:
> On Wed, 25 Mar 2020 15:31:08 +0100
> Giulio Benetti <giulio.benetti at benettiengineering.com> wrote:
> 
>> +diff --git a/nss/lib/freebl/Makefile b/nss/lib/freebl/Makefile
>> +index dc502f309..c7eb9c205 100644
>> +--- a/nss/lib/freebl/Makefile
>> ++++ b/nss/lib/freebl/Makefile
>> +@@ -781,7 +781,7 @@ ifeq ($(CPU_ARCH),arm)
>> + USES_SOFTFLOAT_ABI := $(shell $(CC) -o - -E -dM - $(CFLAGS) < /dev/null | grep __SOFTFP__ > /dev/null && echo 1)
>> + $(OBJDIR)/$(PROG_PREFIX)aes-armv8$(OBJ_SUFFIX): CFLAGS += -march=armv8-a -mfpu=crypto-neon-fp-armv8$(if $(USES_SOFTFLOAT_ABI), -mfloat-abi=softfp)
>> + ifndef NSS_DISABLE_GCM_ARM32_NEON
>> +-$(OBJDIR)/$(PROG_PREFIX)gcm-arm32-neon$(OBJ_SUFFIX): CFLAGS += -march=armv7 -mfpu=neon$(if $(USES_SOFTFLOAT_ABI), -mfloat-abi=softfp)
>> ++$(OBJDIR)/$(PROG_PREFIX)gcm-arm32-neon$(OBJ_SUFFIX): CFLAGS += -mfpu=neon$(if $(USES_SOFTFLOAT_ABI), -mfloat-abi=softfp)
> 
> But only ARMv7 can have NEON support support. So removing just
> -march=armv7 doesn't make sense. Actually, the issue is probably that
> they should be using -march=armv7-a, because -march=armv7 is the common
> subset of ARMv7-A, ARMv7-M and ARMv7-R, which I suppose doesn't include
> NEON.
> 
> So while the patch work, its description is misleading at best,

yes, sorry

> and
> having -march=armv7-a here is probably the right fix.

But build failure we have here:
https://gitlab.com/buildroot.org/buildroot/-/jobs/483712746
is on cortex-A53 in 32-bit mode, so NEON is present and we are on arm,
but it is an armv8, not armv7, so we can't pass -march=armv7 to gcc.

Also they have added -march=armv7 to fix an armv6 build failure:
https://bugzilla.mozilla.org/show_bug.cgi?id=1612177
but how can that work on armv6? Probably it builds because -march=armv7 
emits NEON macro in some way, but it shouldn't run correctly.

So basically with a better description I think my patch is correct.
The problem here is the case on armv8(32-bits mode) with -march=armv7

Another way is to avoid building gcm-arm32-neon.c on armv8 in 32 bits mode.

What do you think?

Kind regards
-- 
Giulio Benetti
Benetti Engineering sas