MPI inline assembly doesn't compile for PIC on i386 with older gcc

[simonbutcher]

Note: This is just a template, so feel free to use/remove the unnecessary things

Description

• Type: Bug
• Priority: Minor

---

Bug

mbed TLS build:
Version: mbedtls-2.12.0

With older versions of gcc, Mbed TLS is unable to compile in very specific circumstances:-

• if optimisation is enabled (e.g. -O1 or -O2)
• for an i386 target
• with a version of gcc at least earlier than v6
• when position independent code is enable for a shared library (e.g. -fPIC and -shared)

The problem is that the inline assembly for i386 specifies a lot of registers to clobber, and older versions of gcc can't compile with so few registers.

To reproduce the fault:-

• use an older version of Debian i386. I used Jessue 8.5 (which at time of writing is old but still supported). It supplies gcc version 4.9.2.
• download the Mbed TLS source code and execute the command:

gcc -O2 -fPIC -Iinclude -shared library/bignum.c

This will result in the following error:

In file included from library/bignum.c:47:0:
library/bignum.c: In function ‘mpi_mul_hlp’:
include/mbedtls/bn_mul.h:46:13: error: PIC register clobbered by ‘ebx’ in ‘asm’
 #define asm __asm
             ^
include/mbedtls/bn_mul.h:62:5: note: in expansion of macro ‘asm’
     asm(                                    \
     ^
library/bignum.c:1142:9: note: in expansion of macro ‘MULADDC_INIT’
         MULADDC_INIT
         ^
include/mbedtls/bn_mul.h:46:13: error: PIC register clobbered by ‘ebx’ in ‘asm’
 #define asm __asm
             ^

Workarounds

There’s a couple of ways around this:

• the easiest option is probably going to be to update the compiler. The code will work with gcc 7 and gcc 6 with no issue. There's also no issue with clang, so swapping to clang may fix this.

• you can specify to turn off inline assembly. This can be done by commenting out the line ’#define MBEDTLS_HAVE_ASM in the Mbed TLS configuration file, found in include/mbedlts/config.h. That slows down some of the asymmetric crypto operations slower, but at least it builds.

[ciarmcom]

ARM Internal Ref: IOTSSL-2461

[wyattoday]

We encountered this bug and worked around this problem by modifying include/mbedtls/bn_mul.h to not use the i386 "big num" assembly when -fPIC is used during the compilation. So, modifying this...

#if defined(__i386__) && defined(__OPTIMIZE__)

#define MULADDC_INIT                        \
    asm(                                    \
....

To this:

#if defined(__i386__) && defined(__OPTIMIZE__) && !defined(__PIC__)

#define MULADDC_INIT                        \
    asm(                                    \
....

Note that __PIC__ is defined if -fPIC is used during compilation.

This solution for us was preferable to the other options (using a newer compiler, dropping support for x86, not using position independent code, disabling ASM throughout mbedTLS, etc.)

[mckaygerhard]

we need #1986 merged please

[gilles-peskine-arm]

@mckaygerhard Can you please make a rebased version of it for the three supported branches (development, development_2.x, mbedtls-2.16)?

[mckaygerhard]

i can help with that.. for the weekend if you guys wait for me! thanks to accept the proposal

[mckaygerhard]

I forgot about this, I was too busy, this patch must be apply .. yet many devices still use gcc 4.X for example embedded... this patch should still be supported because linux kernels like 4.X are still in use... for devices that can't use 5.X as an example, i cannot rebase, cos i used older devices (have a lot of those with large life and i will not change cos represent a cost)

i will try again for, next weekend if there's no response

[tom-cosgrove-arm]

Note that the supported branches are now development and mbedtls-2.28