On 30/09/2021 02.18, Nick Desaulniers wrote:
> On Wed, Sep 29, 2021 at 4:28 PM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>>
> Though for the defconfig case...somehow the cost is more than with the
> sanitizers...
>
> arch/x86/mm/amdtopology.c:157:7: remark: '__nodes_weight' not inlined
> into 'amd_numa_init' because too costly to inline (cost=930,
> threshold=45) [-Rpass-missed=inline]
> if (!nodes_weight(numa_nodes_parsed))
> ^
>
> Looking at the output of `make LLVM=1 -j72
> arch/x86/mm/amdtopology.ll`, @__nodes_weight is just some inline asm
> (.altinstructions). I wonder if I need to teach the cost model about
> `asm inline`...
Remind me, does clang understand 'asm inline("foo")'? Regardless, it
seems that the
asm (ALTERNATIVE("call __sw_hweight32", ...
asm (ALTERNATIVE("call __sw_hweight64", ...
in arch/x86/include/asm/arch_hweight.h could/should be made asm_inline
at least for gcc's sake.
Somewhat related: I really think we should remove __cold from the
definition of __init: It hurts boot time (on a simple board with quite
reproducible boot timing I measured 1-3% some time ago), and it is
likely at least partially responsible for the never-ending tsunami of
functions-that-obviously-should-have-been-inlined(TM) but were not
because the caller is being optimized for size. Whatever small cost in
extra .text is reclaimed after init - and those who are concerned about
the size of the kernel image itself probably build with
CONFIG_OPTIMIZE_FOR_SIZE=y, and I see no change in such an image whether
__init includes __cold or not.
Rasmus
