Re: [PATCH] arm64: Fix multiple 'asm-operand-widths' warnings

From: Mark Rutland
Date: Wed May 03 2017 - 06:52:36 EST


On Tue, May 02, 2017 at 11:52:12AM -0700, Matthias Kaehlcke wrote:
> El Tue, May 02, 2017 at 06:29:48PM +0100 Mark Rutland ha dit:
> > On Mon, May 01, 2017 at 02:26:22PM -0700, Matthias Kaehlcke wrote:
> > > diff --git a/arch/arm64/include/asm/uaccess.h b/arch/arm64/include/asm/uaccess.h
> > > index 5308d696311b..7db143689694 100644
> > > --- a/arch/arm64/include/asm/uaccess.h
> > > +++ b/arch/arm64/include/asm/uaccess.h
> > > @@ -302,7 +302,7 @@ do { \
> > > " .previous\n" \
> > > _ASM_EXTABLE(1b, 3b) \
> > > : "+r" (err) \
> > > - : "r" (x), "r" (addr), "i" (-EFAULT))
> > > + : "r" ((__u64)x), "r" (addr), "i" (-EFAULT))
> > >
> >
> > For reference, do you have the warning for this case to hand?
> >
> > In __put_user_err() we make __pu_val the same type as *ptr, then we
> > switch on sizeof(*ptr), and pass __pu_val to __put_user_asm(), as x.
> > For cases 1, 2, and 4, we use "%w" as the register template.
> >
> > So I can't see why we'd need this cast in __put_user_err().
> >
> > I must be missing something.
>
> This is one of many instances:
>
> ./include/linux/pagemap.h:554:10: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
> return __put_user(0, end);
> ^
> ./arch/arm64/include/asm/uaccess.h:338:2: note: expanded from macro '__put_user'
> __put_user_err((x), (ptr), __pu_err); \
> ^
> ./arch/arm64/include/asm/uaccess.h:326:38: note: expanded from macro '__put_user_err'
> __put_user_asm("str", "sttr", "%", __pu_val, (ptr), \
> ^
> ./include/linux/pagemap.h:554:10: note: use constraint modifier "w"
> ./arch/arm64/include/asm/uaccess.h:338:2: note: expanded from macro '__put_user'
> __put_user_err((x), (ptr), __pu_err); \
> ^
> ./arch/arm64/include/asm/uaccess.h:326:34: note: expanded from macro '__put_user_err'
> __put_user_asm("str", "sttr", "%", __pu_val, (ptr), \
> ^

Thanks for the log above!

> 'end' is a char pointer, it is not clear to me why we would end up in
> the width == 8 branch.

Indeed.

I took a look, and I think the issue is that clang instantiates the
assembly in all cases, producing the warning, *then* optimizes away the
unreachable cases.

If you ask clang to build the following:

----
#define __put_user(val, ptr) \
do { \
__typeof__(*(ptr)) __pu_val = (val); \
switch (sizeof(*(ptr))) { \
case 1: \
asm volatile ("strb %w1, %0" \
: "+Q" (*(ptr)) : "r" (__pu_val)); \
break; \
case 2: \
asm volatile ("strh %1, %0" \
: "+Q" (*(ptr)) : "r" (__pu_val)); \
break; \
case 4: \
asm volatile ("str %1, %0" \
: "+Q" (*(ptr)) : "r" (__pu_val)); \
break; \
case 8: \
asm volatile ("str %1, %0" \
: "+Q" (*(ptr)) : "r" (__pu_val)); \
break; \
} \
} while (0)

void put_char(char in, char *ptr)
{
__put_user(in, ptr);
}
----

It complains for all of the unmatched cases:

----
size-switch.c:26:2: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
__put_user(in, ptr);
^
size-switch.c:11:28: note: expanded from macro '__put_user'
: "+Q" (*(ptr)) : "r" (__pu_val)); \
^
size-switch.c:26:2: note: use constraint modifier "w"
size-switch.c:10:23: note: expanded from macro '__put_user'
asm volatile ("strh %1, %0" \
^
size-switch.c:26:2: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
__put_user(in, ptr);
^
size-switch.c:15:28: note: expanded from macro '__put_user'
: "+Q" (*(ptr)) : "r" (__pu_val)); \
^
size-switch.c:26:2: note: use constraint modifier "w"
size-switch.c:14:22: note: expanded from macro '__put_user'
asm volatile ("str %1, %0" \
^
size-switch.c:26:2: warning: value size does not match register size specified by the constraint and modifier [-Wasm-operand-widths]
__put_user(in, ptr);
^
size-switch.c:19:28: note: expanded from macro '__put_user'
: "+Q" (*(ptr)) : "r" (__pu_val)); \
^
size-switch.c:26:2: note: use constraint modifier "w"
size-switch.c:18:22: note: expanded from macro '__put_user'
asm volatile ("str %1, %0" \
^
3 warnings generated.
----

AFAICT, in all other cases where we switch(sizeof(...)), we (will) use
an explicit cast on the parameter, which placates clang.

I think the best option is to get rid of __pu_val, and have an explicit
cast of x in each case of the switch statement. I'll add that to my asm
fixups series, with your Reported-by.

Thanks,
Mark.