Re: randconfig build error with next-20141204, in drivers/pwm

[Thierry Reding]

On Thu, Dec 18, 2014 at 10:44:44AM +0100, Thierry Reding wrote:
> On Thu, Dec 04, 2014 at 09:10:55AM -0700, Jim Davis wrote:
> > Building with the attached random configuration file,
> > 
> > ERROR: "____ilog2_NaN" [drivers/pwm/pwm-atmel-hlcdc.ko] undefined!
> 
> This took a while to figure out. The attached patch fixes this build
> failure, though the driver should probably be fixed to avoid division by
> zero, just in case. Adding Boris for visibility.
> 
> Thierry

> From 7933af1d2e5f3941d934eec88f32f5547ee218c3 Mon Sep 17 00:00:00 2001
> From: Thierry Reding <thierry.reding@xxxxxxxxx>
> Date: Thu, 18 Dec 2014 10:09:42 +0100
> Subject: [PATCH] pwm: atmel-hlcdc: Depend on HAVE_CLK
> 
> The include/linux/clk.h header defines dummy implementations for the
> various clk_*() functions if HAVE_CLK is not selected to improve build
> coverage in randconfig builds.
> 
> The dummy implementation of clk_get_rate() returns 0, which causes the
> Atmel HLCDC PWM driver's atmel_hlcdc_pwm_config() implementation to end
> up calling:
> 
> 	do_div(clk_period_ns, 0)
> 
> On x86, do_div(n, base) will end up evaluating to this:
> 
> 	n >>= ilog2(base)
> 
> with base = 0, the implementation of ilog2() will call ____ilog2_NaN(),
> which is purposely undefined and results in a linker failure:
> 
> 	ERROR: "____ilog2_NaN" [drivers/pwm/pwm-atmel-hlcdc.ko] undefined!
> 
> The implementation of do_div() checks that base is a power of 2 before
> calling ilog2(). The compiler doesn't optimize this away, presumably
> because is_power_of_2() is an inline function and the compiler doesn't
> or can't inspect it closely enough. ilog2() being a macro it still ends
> up generating the ____ilog2_NaN() because of the constant 0.

If I turn is_power_of_2() into a macro, then this build failure also
goes away. I suppose the reason is that now the do_div() evaluates such
that the branch that would reference the undefined symbol can be
discarded.

Still I think allowing that branch to remain will cause the linker
failure on division by (constant) zero, which has some value, too.

Thierry

Attachment: pgpIMYNdqfj8m.pgp
Description: PGP signature