Re: [GIT PULL] optimize 64-by-32 ddivision for constant divisors on 32-bit machines

[Nicolas Pitre]

On Sun, 22 Nov 2015, Arnd Bergmann wrote:

> On Monday 16 November 2015 20:20:38 you wrote:
> > Arnd,
> > 
> > Please pull the following branch:
> > 
> > 	git://git.linaro.org/people/nicolas.pitre/linux div64
> > 
> > This contains those patches I've initially posted here:
> > 
> > 	https://lkml.org/lkml/2015/11/2/715
> > 
> > Only changes to those posted patches are cosmetic improvements such as 
> > the use of ilog2() replacing the custom __div64_ffs(). Exposure in 
> > linux-next would be a good thing.
> > 
> > I also included fixes for a couple do_div() misuses that an allyesconfig 
> > build turned up after switching ARM to the generic do_div() code.  
> > Those patches have been posted separately and addressed to relevant 
> > maintainers. They are included here until/unless those maintainers 
> > include those patches in their tree.
> > 
> > Original cover letter:
> > 
> > This is a generalization of the optimization I produced for ARM a decade
> > ago to turn constant divisors into a multiplication by the divisor
> > reciprocal. Turns out that after all those years gcc is still not
> > optimizing things on its own for that case.
> > 
> > This has important performance benefits as discussed in this thread:
> > 
> > 	https://lkml.org/lkml/2015/10/28/851
> > 
> > This series brings the formerly ARM-only optimization to all 32-bit
> > architectures using C code by default.  The possibility for the actual
> > multiplication to be implemented in assembly is provided in order to get
> > optimal code.  The ARM version can be used as an example implementation
> > for other interested architectures to implement.
> 
> I'm now getting a build regressing with the attached randconfig configuration,
> when compiling drivers/net/wireless/iwlegacy/common.o:
> 
> drivers/built-in.o: In function `il_send_rxon_timing':
> :(.text+0xbbac80): undefined reference to `__aeabi_uldivmod'
> :(.text+0xbbac9c): undefined reference to `__aeabi_uldivmod'
> :(.text+0xbbacdc): undefined reference to `__aeabi_uldivmod'
> :(.text+0xbbadc8): undefined reference to `__aeabi_uldivmod'
> :(.text+0xbbadf8): undefined reference to `__aeabi_uldivmod'
> :(.text+0xbbae3c): more undefined references to `__aeabi_uldivmod' follow
> drivers/built-in.o: In function `il_send_rxon_timing':
> :(.text+0xbbb11c): undefined reference to `____ilog2_NaN'

This looks like some gcc bug from a few years ago.

> I've verified that this goes away if I turn off CONFIG_PROFILE_ALL_BRANCHES,
> and it only shows up with gcc-5.0 through gcc-5.2.1, but not 4.9.3.
> 
> Aside from those, I have not been able to reduce the failure scenario.
> 
> Also, I have only tested on ARM32, no idea if this shows up elsewhere.

What if you revert "ARM: asm/div64.h: adjust to generic codde" and 
recompile?


Nicolas
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/