Hi,

On 08/05/2020 06:40 PM, Segher Boessenkool wrote:

Hi!

On Wed, Aug 05, 2020 at 04:40:16PM +0000, Christophe Leroy wrote:

quoted

quoted

It cannot optimise it because it does not know shift < 32.  The code
below is incorrect for shift equal to 32, fwiw.

Is there a way to tell it ?

Sure, for example the &31 should work (but it doesn't, with the GCC
version you used -- which version is that?)

GCC 10.1

quoted

quoted

What does the compiler do for just

static __always_inline u64 vdso_shift_ns(u64 ns, unsigned long shift)
	return ns >> (shift & 31);
}

Worse:

I cannot make heads or tails of all that branch spaghetti, sorry.

quoted

  73c:	55 8c 06 fe 	clrlwi  r12,r12,27
  740:	7f c8 f0 14 	addc    r30,r8,r30
  744:	7c c6 4a 14 	add     r6,r6,r9
  748:	7c c6 e1 14 	adde    r6,r6,r28
  74c:	34 6c ff e0 	addic.  r3,r12,-32
  750:	41 80 00 70 	blt     7c0 <__c_kernel_clock_gettime+0x114>

This branch is always true.  Hrm.

As a standalone function:

With your suggestion:

000006ac <vdso_shift_ns>:
  6ac:	54 a5 06 fe 	clrlwi  r5,r5,27
  6b0:	35 25 ff e0 	addic.  r9,r5,-32
  6b4:	41 80 00 10 	blt     6c4 <vdso_shift_ns+0x18>
  6b8:	7c 64 4c 30 	srw     r4,r3,r9
  6bc:	38 60 00 00 	li      r3,0
  6c0:	4e 80 00 20 	blr
  6c4:	54 69 08 3c 	rlwinm  r9,r3,1,0,30
  6c8:	21 45 00 1f 	subfic  r10,r5,31
  6cc:	7c 84 2c 30 	srw     r4,r4,r5
  6d0:	7d 29 50 30 	slw     r9,r9,r10
  6d4:	7c 63 2c 30 	srw     r3,r3,r5
  6d8:	7d 24 23 78 	or      r4,r9,r4
  6dc:	4e 80 00 20 	blr


With the version as is in my series:

000006ac <vdso_shift_ns>:
  6ac:	21 25 00 20 	subfic  r9,r5,32
  6b0:	7c 69 48 30 	slw     r9,r3,r9
  6b4:	7c 84 2c 30 	srw     r4,r4,r5
  6b8:	7d 24 23 78 	or      r4,r9,r4
  6bc:	7c 63 2c 30 	srw     r3,r3,r5
  6c0:	4e 80 00 20 	blr


Christophe