Re: [RFC] proof-of-concept: warning for a bit comparison that is always true/false

[Jeff Law <law at redhat dot com>]

On 04/16/14 09:27, Daniel Marjamäki wrote:
> Hello!
>
> I am new to GCC.
>
> I want to add a warning to GCC when bit comparison is always true/false.
>
> Example:
>
>      if ((x&4)==0) {}  // <- no warning
>
>      if ((x&4)==4) {}  // <- no warning
>
>      if ((x&4)==5) {}  // <- warn!
>
> When this warning is triggered, the most common cause is that somebody
> made a mistake when using bitmasks.
>
> I attach a proof-of-concept patch. I would like comments.
>
> The patch needs some cleanup before it's applied.. I would like it to
> handle at least != also and not just ==. And I would like it to be
> less strict about where integer constants are located.
>
> I wonder where I should put this code. Is gcc/c/c-typeck.c a good file
> to put this in? Should I put it in somewhere else?
>
> What warning flags should be used to enable this? Is some
> -Wcondition-bitop a good idea? Can this be added by -Wall?
>
> I wrote this check for Cppcheck years ago. In my experience this
> warning has a good signal/noise ratio.
I'd actually do this down at the gimple level.  You'll have an SSA graph 
you can use to identify the masking and verify its producing a single 
bit result.  You'll also have canonicalized comparisons, so there'll be 
fewer things to test.  Depending on exactly where you put the 
optimization, you may also see more constants on the RHS due to propagation.


A completely different approach would be to have VRP identify objects 
which have values that are only powers of 2.  Once such a value is in 
the lattice, you can identify and warn/optimize when they're compared 
against values which aren't powers of 2.

Jeff
> Best regards,
> Daniel Marjamäki