[David Michael] [PATCH libgcypt 2/2] cipher/poly1305: Fix 32-bit x86 compilation

[David Michael]

On Fri, Jan 22, 2021 at 10:32 AM Jussi Kivilinna <jussi.kivilinna at iki.fi> wrote:
> Hello,
>
> On 22.1.2021 13.21, Werner Koch via Gcrypt-devel wrote:
> > ForwardedMessage.eml / David Michael <fedora.dm0 at gmail.com>:
> >
> > * cipher/poly1305.c [HAVE_COMPATIBLE_GCC_ARM_PLATFORM_AS]: Also
> > conditionalize on whether __arm__ is defined.
> >
> > --
> >
> > When building for i686, configure detects that the assembler can
> > use the different architectures, so it defined everything under the
> > HAVE_COMPATIBLE_GCC_ARM_PLATFORM_AS conditional block.  Since that
> > block is first and the following x86 block only defines UMUL_ADD_32
> > if it's not already defined, the lto-wrapper failed during linking
> > with a pile of "no such instruction: umlal ..." errors.  Gating on
> > __arm__ prevents that initial defintion and fixes the errors.
> >
>
> HAVE_COMPATIBLE_GCC_ARM_PLATFORM_AS should not be defined on x86/i686
> in first place.

Using i686 GCC has these in config.log:

#define HAVE_COMPATIBLE_GCC_ARM_PLATFORM_AS 1
#define HAVE_COMPATIBLE_GCC_AARCH64_PLATFORM_AS 1
#define HAVE_COMPATIBLE_GCC_AMD64_PLATFORM_AS 1

> Problem must be in 'configure.ac' in the HAVE_COMPATIBLE_GCC_ARM_PLATFORM_AS
> check. Does this issue happen with "clang -flto"?

If I set CC to i686 clang, it first fails because it doesn't undestand
this line in jitterentropy-base.c

#pragma GCC optimize ("O0")

If I use -O0 in CFLAGS, then it fails due to the ARM instructions
while linking libgcrypt.so again.  It has the same HAVE_COMPATIBLE_GCC
definitions in config.log as with GCC.

<inline asm>:1:2: error: invalid instruction mnemonic 'umlal'
        umlal %ecx, %eax, %edx, %esi
        ^~~~~
LLVM ERROR: Error parsing inline asm

Thanks.

David