Re: [PATCH] s390/test_unwind: use raw opcode instead of invalid instruction

From: Christian Borntraeger
Date: Fri Nov 19 2021 - 04:39:34 EST

Am 19.11.21 um 02:10 schrieb Nick Desaulniers:
On Wed, Nov 17, 2021 at 9:48 AM Ilie Halip <ilie.halip@xxxxxxxxx> wrote:

Building with clang & LLVM_IAS=1 leads to an error:
arch/s390/lib/test_unwind.c:179:4: error: invalid register pair
" mvcl %%r1,%%r1\n"

The test creates an invalid instruction that would trap at runtime, but the
LLVM inline assembler tries to validate it at compile time too.

Use the raw instruction opcode instead.

Reported-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
Signed-off-by: Ilie Halip <ilie.halip@xxxxxxxxx>

Ilie, thanks for the patch!

So if I understand
that `e,` prefix is for 16B opcodes?

e is an instruction format as specified by the architecture.
without any parameters.
Normally RR would be the right thing for MVCL, but since
we try to build an invalid opcode without the assembler
noticing (ab)using e seem like a safer approach.

LGTM, thanks again.
Suggested-by: Ulrich Weigand <Ulrich.Weigand@xxxxxxxxxx>
Reviewed-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>

added those and added my RB. applied to the s390 tree. Thanks

I triple checked that GAS, clang, and GNU objdump are in agreement in
terms of encoding here.

arch/s390/lib/test_unwind.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/s390/lib/test_unwind.c b/arch/s390/lib/test_unwind.c
index cfc5f5557c06..d342bc884b94 100644
--- a/arch/s390/lib/test_unwind.c
+++ b/arch/s390/lib/test_unwind.c
@@ -176,7 +176,7 @@ static noinline int unwindme_func4(struct unwindme *u)
* trigger specification exception
asm volatile(
- " mvcl %%r1,%%r1\n"
+ " .insn e,0x0e11\n" /* mvcl %%r1,%%r1" */
"0: nopr %%r7\n"
EX_TABLE(0b, 0b)