Re: [PATCH next 11/14] bit: Strengthen compile-time tests in GENMASK() and BIT()

[Vincent Mailhol]

On 21/01/2026 at 15:57, david.laight.linux@xxxxxxxxx wrote:
> From: David Laight <david.laight.linux@xxxxxxxxx>
> 
> The current checks in GENMASK/BIT (eg reversed high/low) only work
> for 'integer constant expressions' not 'compile-time constants'.
> This is true for const_true() and -Wshift-count-overflow/negative.
> While compile-time constants may be unusual, they can happen through
> function inlining.

Did those new checks actually found any real problem in the code? This
adds a lot of complexity so I am not sure whether this is a winning
trade-off.

> This isn't too bad with gcc, but if clang detects a negative/over-large
> shift it treats it as 'undefined behaviour' and silently discards all
> code that would use the result, so:
> int f(u32 x) {int n = 32; return x >> n; }
> generates a function that just contains a 'return' instruction.
> If 'n' was a variable that happened to be 32, most modern cpu mask
> the count - so would return 'x', some might return 0.

But then, you only solve that shift problem for GENMASK() and
BIT(). Any other usage of the left/right shifts are not diagnosed
unless your check get copy pasted all over the place.

I think that such a check belongs to a static analyzer. Speaking of
which:

	$ cat test.c
	typedef unsigned int u32;
	static int f(u32 x) {int n = 32; return x >> n; }

	$ sparse test.c
	test.c:2:46: warning: shift too big (32) for type unsigned int$ cat test.c

So here, I would rather keep relying on sparse rather that introducing
the W=c logic and all that macro complexity.


Yours sincerely,
Vincent Mailhol