On Mon, Jan 20, 2020 at 06:08:23PM +0100, Christophe Leroy wrote:

Not easy I think.

First we have the unavoidable ASM entry function that can't be dropped 
because of the CR[SO] bit the set on error or clear on no error and that 
can't be done in C.

Yup.

In our ASM VDSO, fixed shifts are used, while in generic C VDSO, shifts 
are generic and read from the VDSO data.

Does that cost more than just a few cycles?

And there is still some funny code generated by GCC (8.1), like:

 620:	7d 29 3c 30 	srw     r9,r9,r7
 624:	21 87 00 20 	subfic  r12,r7,32
 628:	7d 07 3c 31 	srw.    r7,r8,r7
 62c:	7d 08 60 30 	slw     r8,r8,r12
 630:	7d 0b 4b 78 	or      r11,r8,r9

(This can be done cheaper for fixed shifts, you can use rlwimi then).

 634:	39 40 00 00 	li      r10,0
 638:	40 82 00 84 	bne     6bc <__c_kernel_clock_gettime+0x114>
 63c:	81 23 00 24 	lwz     r9,36(r3)
 640:	81 05 00 00 	lwz     r8,0(r5)
...
 6bc:	7d 69 5b 78 	mr      r9,r11
 6c0:	7c ea 3b 78 	mr      r10,r7
 6c4:	7d 2b 4b 78 	mr      r11,r9
 6c8:	4b ff ff 74 	b       63c <__c_kernel_clock_gettime+0x94>

This branch to 6bc is totally useless:
- copying r11 into r9 is pointless as r9 is overwritten in 63c
- copying back r9 into r11 is pointless as r11 has not been modified 
inbetween.

Yeah, huh, how did that happen.

- loading r10 with 0 then overwritting r10 with r7 when r7 is not 0 is 
pointless as well, could have directly put the result of srw. in r10.

This may be harder to make the compiler do.

But the r9/r11 thing suggests you are preventing optimisation somewhere,
maybe with some asm?  Do you have some small testcase I can compile?


Segher

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel