Re: [PATCH v4 8/8] lib: test_bitmap: add compile-time optimization/evaluations assertions

[Alexander Lobakin]

From: Alexander Lobakin <alexandr.lobakin@xxxxxxxxx>
Date: Tue, 21 Jun 2022 21:15:53 +0200

> Add a function to the bitmap test suite, which will ensure that
> compilers are able to evaluate operations performed by the
> bitops/bitmap helpers to compile-time constants when all of the
> arguments are compile-time constants as well, or trigger a build
> bug otherwise. This should work on all architectures and all the
> optimization levels supported by Kbuild.
> The function doesn't perform any runtime tests and gets optimized
> out to nothing after passing the build assertions.
> 
> Suggested-by: Yury Norov <yury.norov@xxxxxxxxx>
> Signed-off-by: Alexander Lobakin <alexandr.lobakin@xxxxxxxxx>
> Reviewed-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> ---
>  lib/test_bitmap.c | 45 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 45 insertions(+)
> 
> diff --git a/lib/test_bitmap.c b/lib/test_bitmap.c
> index d5923a640457..3a7b09b82794 100644
> --- a/lib/test_bitmap.c
> +++ b/lib/test_bitmap.c
> @@ -869,6 +869,50 @@ static void __init test_bitmap_print_buf(void)
>  	}
>  }
>  
> +static void __init test_bitmap_const_eval(void)
> +{
> +	DECLARE_BITMAP(bitmap, BITS_PER_LONG);
> +	unsigned long initvar = BIT(2);
> +	unsigned long bitopvar = 0;
> +	unsigned long var = 0;
> +	int res;
> +
> +	/*
> +	 * Compilers must be able to optimize all of those to compile-time
> +	 * constants on any supported optimization level (-O2, -Os) and any
> +	 * architecture. Otherwise, trigger a build bug.
> +	 * The whole function gets optimized out then, there's nothing to do
> +	 * in runtime.
> +	 */
> +
> +	/* Equals to `unsigned long bitmap[1] = { BIT(5), }` */
> +	bitmap_clear(bitmap, 0, BITS_PER_LONG);
> +	if (!test_bit(7, bitmap))
> +		bitmap_set(bitmap, 5, 1);

So for now, when building for s390, Clang (up to the latest Git
snapshot) generates some incorrect code here.
It does expand both test_bit() and bitmap_set() to const_test_bit()
and const___set_bit(), but at the same time thinks that starting
from this point, @bitmap and @bitopvar (???) are *not* constants
and fails the assertions below, which is not true and weird.
Any other architecture + compiler couples work fine, including
s390 on GCC.
So would it be acceptable for now to do:

	/* Equals to `unsigned long bitmap[1] = { BIT(5), }` */
	bitmap_clear(bitmap, 0, BITS_PER_LONG);
	/*
	 * Some comment saying that this is currently broken
	 * on s390 + Clang
	 */
#if !(defined(__s390__) && defined(__clang__))
	if (!test_bit(7, bitmap))
		bitmap_set(bitmap, 5, 1);
#endif

	/* Equals to `unsigned long bitopvar = BIT(20)` */
	__change_bit(31, &bitopvar);
	bitmap_shift_right(&bitopvar, &bitopvar, 11, BITS_PER_LONG);

[...]

or there could be any better solutions?

(+Cc LLVM folks)

> +
> +	/* Equals to `unsigned long bitopvar = BIT(20)` */
> +	__change_bit(31, &bitopvar);
> +	bitmap_shift_right(&bitopvar, &bitopvar, 11, BITS_PER_LONG);
> +
> +	/* Equals to `unsigned long var = BIT(25)` */
> +	var |= BIT(25);
> +	if (var & BIT(0))
> +		var ^= GENMASK(9, 6);
> +
> +	/* __const_hweight<32|64>(BIT(5)) == 1 */
> +	res = bitmap_weight(bitmap, 20);
> +	BUILD_BUG_ON(!__builtin_constant_p(res));
> +
> +	/* !(BIT(31) & BIT(18)) == 1 */
> +	res = !test_bit(18, &bitopvar);
> +	BUILD_BUG_ON(!__builtin_constant_p(res));
> +
> +	/* BIT(2) & GENMASK(14, 8) == 0 */
> +	BUILD_BUG_ON(!__builtin_constant_p(initvar & GENMASK(14, 8)));
> +	/* ~BIT(25) */
> +	BUILD_BUG_ON(!__builtin_constant_p(~var));
> +}
> +
>  static void __init selftest(void)
>  {
>  	test_zero_clear();
> @@ -884,6 +928,7 @@ static void __init selftest(void)
>  	test_for_each_set_clump8();
>  	test_bitmap_cut();
>  	test_bitmap_print_buf();
> +	test_bitmap_const_eval();
>  }
>  
>  KSTM_MODULE_LOADERS(test_bitmap);
> -- 
> 2.36.1

Thanks,
Olek