I futzed with porting my other os to clang and this came up. It seems that
GAS is more tolerant of the form of the instruction and how the suffixes
work. It should accept the version that clang is happy with. The clang
assembler only accepts one form of it.
Similar things happen with all the cp14 registers for fpu control bits that
now have official accessor instruction aliases. Clang only accepts the
newest definition of the mnemonics, and doesn't apparently have any code to
parse mrc/mcr !p15. ('vmsr fpexc' and friends were the offenders).
All this aside, if it's a problem you can tell clang to not use the built
in assembler and punch through to gas, which I'm assuming you'll still need
for now since binutils's ld is still more mature for this sort of thing,
versus llvm's lld.
On Sun, Nov 8, 2015 at 1:10 AM, François Revol <revol@xxxxxxx> wrote:
On 08/11/2015 08:53, Adrien Destugues wrote:
On Sat, Nov 07, 2015 at 06:30:49PM +0100, js-haiku-commits@xxxxxxxxxxxwrote:
55de0addd599: arm/arch_string.S: Only use sane insns for Clang
Are they broken on GNU AS? And can't we fix that instead of duplicating
the code?
Another solution is to #define one another.
François.
