Re: [PATCH] arm: Save and restore VFP arguments registers in PLT trampolines

["Joseph S. Myers" <joseph at codesourcery dot com>]

On Wed, 11 Apr 2012, Meador Inge wrote:

> +#if defined(__ARM_NEON__)

That's not the right condition.

The issue is not specific to NEON (__ARM_NEON__), or to the VFP AAPCS 
variant (__ARM_PCS_VFP).  It's easier to detect that there's a problem 
when the VFP AAPCS variant is used, because then the issue can corrupt 
function arguments.  But the PLT may be used to call functions such as 
__aeabi_memcpy whose ABIs are defined to clobber only registers that are 
call-clobbered in the base AAPCS, not call-clobbered registers specific to 
the VFP variant.  So the issue applies to all call-clobbered VFP registers 
regardless of the ABI.  Furthermore, the VFP ABI variant can be used with 
any version of VFP (including the single-precision-only versions, since 
those still have double-precision moves), and does not need NEON (and I 
made sure that GCC uses the same ABI for passing generic vectors when 
using the VFP ABI variant for non-NEON VFP as it does for passing NEON 
vector types with NEON enabled - these are fully compatible).

Of course the issue only applies if something in the dynamic linker 
actually runs code, other than the function being called through the PLT, 
that clobbers such registers.  This can happen when the dynamic linker is 
built with VFP enabled - it may be used internally by the compiler.  There 
may also be cases with audit modules; I'm not expert on those, but the 
x86_64 dynamic linker has code to handle saving and restoring AVX 
registers, based on dynamic detection of AVX (_dl_x86_64_restore_sse, 
RTLD_ENABLE_FOREIGN_CALL, etc.).

Without regard to audit modules, the basic requirement is to save and 
restore d0-d7 if available (defined (__VFP_FP__) && !defined 
(__SOFTFP__)).  And then d16-d31 must also be saved and restored if 
available - but there isn't a preprocessor define for that.  I don't know 
how readily you can check HWCAPs (HWCAP_ARM_VFPv3, HWCAP_ARM_VFPv3D16) 
here.  Audit modules may require some dynamic detection in any case.

-- 
Joseph S. Myers
joseph@codesourcery.com