Re: Build performance regressions originating from min()/max() macros

["Jürgen Groß" <jgross@suse.com>]

On 23.07.24 23:59, Lorenzo Stoakes wrote:
> Arnd reported a significant build slowdown [0], which was bisected to the
> series spanning commit 80fcac55385c ("minmax: relax check to allow
> comparison between unsigned arguments and signed constants") to commit
> 867046cc70277 ("minmax: relax check to allow comparison between unsigned
> arguments and signed constants"), originating from the series "minmax:
> Relax type checks in min() and max()." [1].
> 
> I have reproduced this locally, reverting this series and manually fixing
> up all call sites that invoke min()/max() for a simple x86-64 defconfig (+
> some other debug flags I use for debug kernels, I can provide the .config
> if needed).
> 
> Arnd noted that the arch/x86/xen/setup.c file was particularly problematic,
> taking 15 (!) seconds to pre-process on his machine, so I also enabled
> CONFIG_XEN to test this and obtained performance numbers with this set/not
> set.
> 
> I was able to reproduce this very significant pre-processor time on this
> file, noting that with the series reverted compile time for the file is
> 0.79s, with it in place, it takes 6.90s for a 873.4% slowdown.
> 
> I also checked total build times (32-core intel i9-14900KF box):
> 
> ## With CONFIG_XEN
> 
> ### Reverted minmax code
> 
> make 1588.46s user 92.33s system 2430% cpu 1:09.16 total
> make 1598.57s user 93.49s system 2419% cpu 1:09.94 total
> make 1598.99s user 92.49s system 2419% cpu 1:09.91 total
> 
> ### Not reverted
> 
> make 1639.25s user 96.34s system 2433% cpu 1:11.32 total
> make 1640.34s user 96.01s system 2427% cpu 1:11.54 total
> make 1639.98s user 96.76s system 2436% cpu 1:11.27 total
> 
> ## Without CONFIG_XEN
> 
> ### Reverted minmax code
> 
> make 1524.97s user 89.84s system 2399% cpu 1:07.31 total
> make 1521.01s user 88.99s system 2391% cpu 1:07.32 total
> make 1530.75s user 89.65s system 2389% cpu 1:07.83 total
> 
> ### Not reverted
> 
> make 1570.64s user 94.09s system 2398% cpu 1:09.41 total
> make 1571.25s user 94.36s system 2401% cpu 1:09.36 total
> make 1568.25s user 93.83s system 2396% cpu 1:09.35 total
> 
> Which suggests a worryingly significant slowdown of ~45s with CONFIG_XEN
> enabled and ~35s even without it.
> 
> The underlying problems seems to be very large macro expansions, which Arnd
> noted in the xen case originated from the line:
> 
> extra_pages = min3(EXTRA_MEM_RATIO * min(max_pfn, PFN_DOWN(MAXMEM)),
> 		extra_pages, max_pages - max_pfn);
> 
> And resulted in the generation of 47 MB (!) of pre-processor output.
> 
> It seems a lot of code now relies on the relaxed conditions of the newly
> changed min/max() macros, so the question is - what can we do to address
> these regressions?

I can send a patch to simplify the problematic construct, but OTOH this
will avoid only one particularly bad example.


Juergen