Message-ID: <CACG_h5pDHCp_b=UJ7QZCEDqmJgUdPSaNLR+0sR1Bgc4eCbqEKw@mail.gmail.com>
Date:   Sat, 30 May 2020 14:15:28 +0530
From:   Syed Nayyar Waris <syednwaris@...il.com>
To:     Andy Shevchenko <andy.shevchenko@...il.com>
Cc:     Andy Shevchenko <andriy.shevchenko@...ux.intel.com>,
        Linus Walleij <linus.walleij@...aro.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        William Breathitt Gray <vilhelm.gray@...il.com>,
        Arnd Bergmann <arnd@...db.de>,
        Linux-Arch <linux-arch@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v7 1/4] bitops: Introduce the the for_each_set_clump macro

On Sat, May 30, 2020 at 3:49 AM Andy Shevchenko
<andy.shevchenko@...il.com> wrote:
>
> On Sat, May 30, 2020 at 1:11 AM Syed Nayyar Waris <syednwaris@...il.com> wrote:
> > On Sat, May 30, 2020 at 3:13 AM Andy Shevchenko
> > <andy.shevchenko@...il.com> wrote:
> > > On Fri, May 29, 2020 at 11:07 PM Syed Nayyar Waris <syednwaris@...il.com> wrote:
> > > > On Sat, May 30, 2020 at 12:08 AM Andy Shevchenko
> > > > <andriy.shevchenko@...ux.intel.com> wrote:
> > > > > On Fri, May 29, 2020 at 11:38:18PM +0530, Syed Nayyar Waris wrote:
> > > > > > On Sun, May 24, 2020 at 8:15 PM kbuild test robot <lkp@...el.com> wrote:
> > >
> > > ...
> > >
> > > > > > Taking the example statement (in my patch) where compilation warning
> > > > > > is getting reported:
> > > > > > return (map[index] >> offset) & GENMASK(nbits - 1, 0);
> > > > > >
> > > > > > 'nbits' is of type 'unsigned long'.
> > > > > > In above, the sanity check is comparing '0' with unsigned value. And
> > > > > > unsigned value can't be less than '0' ever, hence the warning.
> > > > > > But this warning will occur whenever there will be '0' as one of the
> > > > > > 'argument' and an unsigned variable as another 'argument' for GENMASK.
> > >
> > > > > Proper fix is to fix GENMASK(), but allowed workaround is to use
> > > > >         (BIT(nbits) - 1)
> > > > > instead.
> > >
> > > > When I used BIT macro (earlier), I had faced a problem. I want to tell
> > > > you about that.
> > > >
> > > > Inside functions 'bitmap_set_value' and 'bitmap_get_value' when nbits (or
> > > > clump size) is BITS_PER_LONG, unexpected calculation happens.
> > > >
> > > > Explanation:
> > > > Actually when nbits (clump size) is 64 (BITS_PER_LONG is 64 on my computer),
> > > > (BIT(nbits) - 1)
> > > > gives a value of zero and when this zero is ANDed with any value, it
> > > > makes it full zero. This is unexpected and incorrect calculation happening.
> > > >
> > > > What actually happens is in the macro expansion of BIT(64), that is 1
> > > > << 64, the '1' overflows from leftmost bit position (most significant
> > > > bit) and re-enters at the rightmost bit position (least significant
> > > > bit), therefore 1 << 64 becomes '0x1', and when another '1' is
> > > > subtracted from this, the final result becomes 0.
> > > >
> > > > Since this macro is being used in both bitmap_get_value and
> > > > bitmap_set_value functions, it will give unexpected results when nbits or clump
> > > > size is BITS_PER_LONG (32 or 64 depending on arch).
> > >
> > > I see, something like
> > > https://elixir.bootlin.com/linux/latest/source/include/linux/dma-mapping.h#L139
> > > should be done.
> > > But yes, let's try to fix GENMASK().
> > >
> > > So, if we modify the following
> > >
> > >   #define GENMASK_INPUT_CHECK(h, l) \
> > >     (BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
> > >     __builtin_constant_p((l) > (h)), (l) > (h), 0)))
> > >
> > > to be
> > >
> > >   #define GENMASK_INPUT_CHECK(h, l) \
> > >     (BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
> > >     __builtin_constant_p((l) > (h)), (l) ? (l) > (h) : 0, 0)))
> > >
> > > would it work?
> >
> > Sorry Andy it is not working. Actually the warning will be thrown,
> > whenever there will be comparison between 'h' and 'l'. If one of them
> > is '0' and the other is unsigned variable.
> > In above, still there is comparison being done between 'h' and 'l', so
> > the warning is getting thrown.
>
> Ah, okay
>
> what about (l) && ((l) > (h)) ?

When I finally changed:
__builtin_constant_p((l) > (h)), (l) > (h), 0)))
to:
__builtin_constant_p((l) && ((l) > (h))), (l) ? (l) > (h) : 0, 0)))

It is still throwing same compilation error at the same location where
comparison is being done between 'l' and 'h'.

Actually the short-circuit logic is not happening. For:
(l) && ((l) > (h))
Even if 'l' is zero, it still proceeds to compare 'l' and 'h' , that
is '((l) > (h))' is checked.
I think it is happening because '__builtin_constant_p' will check the
complete argument: (l) && ((l) > (h)),
'__builtin_constant_p' checks whether the argument is compile time
constant or not, so therefore, it will evaluate the WHOLE argument,
that is (including) the comparison operation.
https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html

I am still investigating more on this. Let me know if you have any suggestions.

>
> > > > William also knows about this issue:
> > > > "This is undefined behavior in the C standard (section 6.5.7 in the N1124)"
> > >
> > > I think it is about 6.5.7.3  here, 1U << 31 (or 63) is okay.
> >
> > Actually for:
> > (BIT(nbits) - 1)
> > When nbits will be BITS_PER_LONG it will be 1U << 32 (or 64). Isn't it ?
> > The expression,
> > BIT(64) - 1
> > can become unexpectedly zero (incorrectly).
>
> Yes, that's why I pointed out to the paragraph. It's about right
> operand to be "great than or equal to" the size of type of left
> operand.
>

Thank You. I understand now. :-)

Regards
Syed Nayyar Waris

Powered by blists - more mailing lists