Re: [PATCH] MIPS: UAPI: Fix unrecognized opcode WSBH/DSBH/DSHD when using MIPS16.
From: Maciej W. Rozycki
Date: Fri Sep 04 2015 - 14:52:52 EST
On Thu, 3 Sep 2015, Yousong Zhou wrote:
> The nomips16 has to be added both as function attribute and assembler
> When only function attribute is specified, the compiler will inline the
> function with -Os optimization. The generated assembly code cannot be
> correctly assembled because ISA mode switch has to be done through jump
This can't be true. The compiler does not intepret the contents of an
inline asm and therefore cannot decide whether to inline a function
containing one or not based on the lone presence or the absence of an
assembly directive within.
> When only ".set nomips16" directive is used, the generated assembly code
> will use MIPS32 code for the inline assembly template and MIPS16 for the
> function return. The compiled binary is invalid:
> 00403100 <__arch_swab16>:
> 403100: 7c0410a0 wsbh v0,a0
> 403104: e820ea31 swc2 $0,-5583(at)
> while correct code should be:
> 00402650 <__arch_swab16>:
> 402650: 7c0410a0 wsbh v0,a0
> 402654: 03e00008 jr ra
> 402658: 3042ffff andi v0,v0,0xffff
It looks to me you're papering something over here -- when you use a
`.set nomips16' directive the assembler will happily switch your
instruction set in the middle of an instruction stream. Consequently if
this function is (for whatever reason; it should not really) inlined in
MIPS16 code, then you'll get a MIPS32 instruction in the middle, which
will obviously be interpreted differently in the MIPS16 execution mode and
is therefore surely a recipe for disaster.
How did you test your change and made the conclusion quoted with your
> Signed-off-by: Yousong Zhou <yszhou4tech@xxxxxxxxx>
> arch/mips/include/uapi/asm/swab.h | 12 +++++++++---
> 1 file changed, 9 insertions(+), 3 deletions(-)
> diff --git a/arch/mips/include/uapi/asm/swab.h b/arch/mips/include/uapi/asm/swab.h
> index 8f2d184..c4ddc4f 100644
> --- a/arch/mips/include/uapi/asm/swab.h
> +++ b/arch/mips/include/uapi/asm/swab.h
> @@ -16,11 +16,13 @@
> #if (defined(__mips_isa_rev) && (__mips_isa_rev >= 2)) || \
> -static inline __attribute_const__ __u16 __arch_swab16(__u16 x)
> +static inline __attribute__((nomips16)) __attribute_const__
> + __u16 __arch_swab16(__u16 x)
> " .set push \n"
> " .set arch=mips32r2 \n"
> + " .set nomips16 \n"
> " wsbh %0, %1 \n"
> " .set pop \n"
> : "=r" (x)
So setting aside the correctness issues discussed above, for MIPS16 code
this has to be put out of line by the compiler, with all the usual
overhead of a function call, causing a performance loss rather than a gain
expected here. Especially as switching the ISA mode requires draining the
whole pipeline so that subsequent instructions are interpreted in the new
execution mode. This is expensive in performance terms.
I'm fairly sure simply disabling these asms (#ifndef __mips16) and
letting the compiler generate the right mask and shift operations from
plain C code will be cheaper in performance terms and possibly cheaper in
code size as well, especially in otherwise leaf functions where an extra
function call will according to the ABI clobber temporaries. So I suggest
going in that direction instead.
So this is a NAK really.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/